Intelligent weeding robot

Abstract

The invention discloses an intelligent weeding robot, and relates to the technical field of weeding robots, the intelligent weeding robot comprises a walking mechanism, a path navigation structure, a weeding structure and a software control and communication module; the walking mechanism adopts an independent driving wheel type assembly, so that stable movement in various terrains is ensured; the path navigation structure integrates a searchlight, a laser radar, a laser obstacle avoidance radar and a scene construction camera, and realizes accurate positioning and dynamic obstacle avoidance through a multi-sensor fusion technology; the weeding structure comprises a binocular intelligent camera, a data processing module and an execution mechanism, the data processing module operates an efficient weed recognition algorithm, and the execution mechanism achieves accurate weeding through a three-axis mechanical arm and a pumping source laser; the software control and communication module coordinates the operation of each component, and the real-time performance of the system is improved. The problems that a traditional weeding method is low in efficiency, pollutes the environment and the like are solved, the method is suitable for a complex farmland environment, and an intelligent and environment-friendly weeding solution is provided for modern agriculture.

Description

Technical Field

[0001] The present invention relates to the technical field of weeding robots, and more particularly, to a robot for intelligent weeding. Background Art

[0002] With the development of modern agriculture towards intelligence and precision, many deficiencies of traditional weeding methods have gradually emerged. Manual weeding relies on a large amount of labor, with low efficiency and high costs, and it is particularly difficult to meet the requirements of large-scale production in vast farmlands. Although chemical weeding is relatively efficient, long-term use of herbicides can lead to soil pollution, ecological imbalance, and the enhancement of weed resistance, posing challenges to the sustainable development of agriculture. In addition, although mechanical weeding equipment can replace manual labor to a certain extent, it generally has problems such as significant damage to crops, insufficient weeding accuracy, and poor adaptability to complex terrains.

[0003] In recent years, the application of intelligent robot technology in the agricultural field has provided new ideas for solving the above problems. Weeding robots based on visual recognition and automatic control have gradually emerged, which can sense the environment through sensors, identify weeds, and perform precise weeding operations. For example, in the prior art, there is already a weed recognition system based on a monocular camera, which uses simple image processing algorithms (such as color segmentation or edge detection) to distinguish weeds from crops. However, the recognition accuracy of such systems is limited under complex lighting conditions, and they cannot accurately obtain the spatial position information of weeds, resulting in poor weeding effects. In addition, some robots use mechanical removal or spraying of herbicides, which, although simple to operate, have obvious deficiencies in eradicating weeds or protecting the environment.

[0004] In terms of navigation technology, existing weeding robots mostly rely on GPS positioning combined with preset paths for movement. However, in the farmland environment, GPS signals may be inaccurate due to occlusion, and they lack real-time obstacle avoidance capabilities, making it difficult to deal with dynamic obstacles or terrain changes. Although lidar and visual SLAM technologies have been applied to the navigation of some robots, their integration level is low and the algorithm complexity is high, making it difficult to run in real time on low-power embedded devices. In addition, there are also limitations in the design of weeding actuators. For example, traditional robotic arms have a small movement range and insufficient flexibility, or the power control of laser weeding systems is inaccurate, resulting in high energy consumption or incomplete weeding.

[0005] In the field of weed recognition algorithms, convolutional neural networks (CNNs) based on deep learning have been used to improve recognition accuracy, such as pre-trained models based on the ImageNet dataset. However, these models have poor adaptability to weed species in agricultural scenarios and do not fully utilize temporal data or the latest advances in object detection (such as the YOLO series). In recent years, the emergence of efficient algorithms such as YOLOv10 and Mamba has made real-time object detection possible, but their application in farmland weed recognition is still in the exploratory stage and lacks deep integration with robot hardware.

[0006] In summary, there is still room for improvement in the existing intelligent weeding robot technology in terms of path navigation accuracy, weed recognition efficiency, flexibility of the execution mechanism, and environmental protection. To address these issues, we urgently need to design a robot for intelligent weeding to solve the above-mentioned problems. Summary of the Invention

[0007] The purpose of the present invention is to solve the problems of the existing technology mentioned in the above background art and provide a robot for intelligent weeding.

[0008] The above object of the present invention is achieved as follows:

[0009] A robot for intelligent weeding, comprising:

[0010] A traveling mechanism for driving the entire robot to travel, the traveling mechanism comprising four independently driven wheel assemblies, each wheel assembly being equipped with a DC brushless motor;

[0011] A path navigation structure for performing path navigation and sending instructions to the traveling mechanism to control its travel, the path navigation structure comprising: a searchlight, a lidar, and a laser obstacle avoidance radar, wherein the searchlight and the lidar are both provided on the top of the housing, the laser obstacle avoidance radar is provided on the front end face of the vehicle frame, and the housing is connected to the vehicle frame by bolts;

[0012] A weeding structure for removing weeds on the ground, the weeding structure comprising:

[0013] A vision detection module, using a binocular intelligent camera, for obtaining the spatial position information of weeds;

[0014] A data processing module, based on a high-performance GPU with a computing power of 12 TFLOPS, running a weed recognition algorithm and performing coordinate calculation and error calibration;

[0015] An execution mechanism, comprising a pump source laser, a high-power beam lens, and a three-axis robotic arm; the binocular intelligent camera, the pump source laser, and the high-power beam lens are all fixed on the output shaft of the three-axis robotic arm, and the three-axis robotic arm is fixedly installed on the vehicle frame;

[0016] A software control and communication module, based on a high-performance GPU, receiving visual signals, running a real-time operating system, and the communication protocols including MQTT and ROS, for coordinating the operation of the traveling mechanism, the path navigation structure, and the weeding structure.

[0017] As a preferred technical solution of the present invention, the path navigation structure constructs a three-dimensional point cloud map of the environment through the lidar and performs SLAM in combination with the scene construction camera. The point cloud generation formula is:

[0018] Px,y,z = {x i , y i , z i | x i = r i · cosθ i · cosφ i , y i = r i · sinθ i ·

[0019] cosφ i , z i = r i · sinφ i};

[0020] Wherein, r i is the distance measured by the lidar; θ i is the horizontal scanning angle, in radians; φ i is the vertical scanning angle, in radians; x i , y i , z i are the point cloud coordinates, in meters; i is the sampling point number; the point cloud density is 1000 points per square meter

[0021] As a preferred technical solution of the present invention, the visual detection module calculates the spatial coordinates of weeds through a binocular intelligent camera, and the coordinate calculation formula is:

[0022]

[0023] Wherein, b is the baseline length, with a value of 120 mm, in millimeters; f is the camera focal length, with a value of 8 mm, in millimeters; d is the parallax of the weed feature points in the left and right images, in pixels; u, v are the pixel coordinates on the image plane, in pixels; X, Y, Z are the spatial coordinates of the weeds, in millimeters; the calculation accuracy of the binocular intelligent camera is ±2 mm, and the parallax d is obtained through the feature point matching algorithm.

[0024] As a preferred technical solution of the present invention, the data processing module uses the CNN based on the CWD30 dataset to fuse the RNN and LSTM algorithms for weed recognition. Among them, the CNN contains 5 convolutional layers and 3 pooling layers, the convolutional kernel size is 3×3 pixels, the stride is 1, and the feature extraction formula is:

[0025] F l = ReLU(W l * I + b l ;

[0026] Wherein, F l is the feature map of the (I) layer, dimensionless; ReLU is the activation function, and:

[0027] ReLUx = max(0, x);

[0028] W l is the convolutional kernel weight matrix, unitless; I is the input image, unit in pixel values (0 - 255); b l is the bias, unitless; * is the convolution operation;

[0029] The RNN and LSTM process time series data, with 128 neurons in the hidden layer and a recognition accuracy of 95%.

[0030] As a preferred technical solution of the present invention, the data processing module uses a weed recognition system based on YOLOv10 and Mamba. The YOLOv10 network includes a Darknet-53 backbone network, with 3 detection boxes, and the loss function is:

[0031]

[0032] where, λ coord is the coordinate loss weight, with a value of 5, unitless;

[0033] λ class : classification loss weight, with a value of 1, unitless;

[0034] S is the grid size, with a value of 13, unitless; B is the number of bounding boxes, with a value of 3, unitless;

[0035] is the object existence indicator, taking values of 0 or 1;

[0036] x i , y i and are the predicted and true coordinates, unit in pixels;

[0037] p i c and are the class probabilities, ranging from 0 - 1;

[0038] The average precision mAP of the optimized Mamba model is greater than 0.92.

[0039] As a preferred technical solution of the present invention, the three-axis robotic arm controls the motion trajectory through inverse kinematics, and the joint angle calculation formula is:

[0040] θ1 = arctan2(y, x)

[0041]

[0042] where, X, Y, Z are the positions of the weed target, unit in millimeters;

[0043] h is the height of the base, with a value of 50 mm and the unit of millimeters;

[0044] L1, L2, and L3 are the lengths of each segment of the robotic arm, which are 200 mm, 200 mm, and 100 mm respectively, and the unit is millimeters;

[0045] θ1, θ2, and θ3 are the joint angles, with the unit of radians; the response time is less than 0.2 s.

[0046] As a preferred technical solution of the present invention, the output power of the pump source laser is dynamically adjusted according to the weed depth, and the calculation formula is:

[0047]

[0048] Among them, P out is the output power, with the unit of watts;

[0049] η is the conversion efficiency, with a value of 0.85 and dimensionless;

[0050] P in is the input power, with a value of 60 W and the unit of watts;

[0051] A lens is the effective area of the lens, with a value of 20 mm 2 , and the unit is square millimeters;

[0052] A spot is the focused spot area, with a value of 0.2 and the unit of square millimeters;

[0053] d is the depth of the weed root, with the unit of millimeters;

[0054] d0 is the reference depth, with a value of 10 mm and the unit of millimeters; the power adjustment range is 20 - 50 W.

[0055] As a preferred technical solution of the present invention, the software control and communication module realizes data transmission through the MQTT protocol, and the message packet size is 256 KB. The calculation formula for the transmission delay is:

[0056]

[0057] T delay is the total delay, with the unit of seconds, and it is required to be less than 0.05 s;

[0058] S data is the data volume, with a value of 256 KB and the unit of bits;

[0059] B is the bandwidth, with a value of 100 Mbps and the unit of bits per second;

[0060] N ops ​The amount of computation, with a value of 10 9 times of computation, with the unit of times;

[0061] F GPU is the computing power, with a value of 12 TFLOPS and the unit of times per second; the calculation result is approximately 0.02 s, meeting the latency requirement.

[0062] As a preferred technical solution of the present invention, the distance threshold D for obstacle detection by the laser obstacle avoidance radar in the path navigation structure th = 0.5 m. When the detected distance D < D th , trigger the A * algorithm to re-plan the path, and the path cost function is:

[0063] fn = gn + hn;

[0064] where f(n) is the total cost, with the unit of meters;

[0065] (g(n)) is the actual distance from the starting point to the current node, with the unit of meters;

[0066] (h(n)) is the Euclidean distance from the current node to the target, with the unit of meters, and the calculation formula is:

[0067] where x n , y n are the coordinates of the current node, with the unit of meters; x g , y g are the coordinates of the target, with the unit of meters; the planning time is less than 1 s, and the resolution is 0.1 m.

[0068] As a preferred technical solution of the present invention, the weeding structure calibrates the coordinate error through the Kalman filter algorithm during the task execution, and the state update formula is:

[0069]

[0070] where is the calibrated state vector, with the unit of millimeters; is the predicted state vector, with the unit of millimeters; K k is the Kalman gain matrix, dimensionless; z k is the observed value, with the unit of millimeters; H is the observation matrix, dimensionless; P k|k-1 is the predicted covariance matrix, with the unit of mm 2 ; R is the measurement noise covariance, with the unit of mm 2 , and the typical value is 4 mm 2 ; the error is reduced to ±1 mm.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] First of all, by integrating an advanced path navigation structure and a walking mechanism, the present invention significantly improves the navigation accuracy and adaptability of the robot in a complex farmland environment. The path navigation structure adopts a multi-sensor fusion technology of lidar, scene construction camera and laser obstacle avoidance radar, and combines SLAM and A* algorithms to be able to construct a high-precision three-dimensional point cloud map in real time and dynamically plan an obstacle avoidance path.

[0073] Secondly, the weeding structure of the present invention combines binocular vision detection, an efficient recognition algorithm and laser weeding technology to achieve accurate recognition and eradication of weeds. The binocular intelligent camera provides a coordinate accuracy of ±2 mm, and the data processing module supports the CNN+RNN+LSTM algorithm and the YOLOv10+Mamba system based on the CWD30 dataset, and can quickly distinguish weeds from crops.

[0074] Finally, the present invention optimizes the system cooperation efficiency through a software control and communication module, enhancing the real-time performance and stability of the operation. Using the MQTT and ROS protocols, the data transmission delay is less than 50 ms. Cooperating with a 12 TFLOPS GPU and the Kalman filtering algorithm, the coordinate error is calibrated to ±1 mm. Compared with the prior art, the present invention improves the weeding efficiency by about 30%, the operation coverage rate reaches 100%, and replaces the traditional method with laser weeding, reducing the damage to the soil and the ecosystem, and providing an intelligent and environmentally friendly solution for modern agriculture. Description of the Drawings

[0075] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0076] Figure 1 is a schematic structural diagram of a robot for intelligent weeding in an embodiment of the present invention;

[0077] Figure 2 is a schematic structural diagram of the weeding structure of a robot for intelligent weeding in an embodiment of the present invention;

[0078] Figure 3 is a logic block diagram of a robot for intelligent weeding in an embodiment of the present invention.

[0079] In the figure: 1, walking mechanism; 2, path navigation structure; 3, weeding structure; 31, three-axis robotic arm; 32, high-power beam lens; 33, pump source laser; 34, binocular intelligent camera. Detailed implementation manners

[0080] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0081] The following will Figures 1 - 3 be described in detail with reference to the accompanying drawings for the specific implementation manners of the present invention.

[0082] The present invention discloses a robot for intelligent weeding, including:

[0083] A traveling mechanism 1 for driving the whole robot to travel. The traveling mechanism includes four independently driven wheel assemblies, and each wheel assembly is equipped with a DC brushless motor;

[0084] A path navigation structure 2 for performing path navigation and sending instructions to the traveling mechanism to control its travel. The path navigation structure includes: a searchlight, a lidar, and a laser obstacle avoidance radar. The searchlight and the lidar are both arranged on the top of the housing, and the laser obstacle avoidance radar is arranged on the front end face of the vehicle frame. The housing is connected to the vehicle frame by bolts;

[0085] A weeding structure 3 for removing weeds on the ground. The weeding structure includes:

[0086] A visual detection module, using a binocular intelligent camera, for obtaining the spatial position information of weeds;

[0087] A data processing module, based on a high-performance GPU with a computing power of 12 TFLOPS, running a weed recognition algorithm and performing coordinate calculation and error calibration;

[0088] An execution mechanism, including a pump source laser 33, a high-power beam lens 32, and a three-axis robotic arm 31; the binocular intelligent camera 34, the pump source laser 33, and the high-power beam lens 32 are all fixed on the output shaft of the three-axis robotic arm 31, and the three-axis robotic arm is fixedly installed on the vehicle frame;

[0089] A software control and communication module, receiving visual signals based on a high-performance GPU, running a real-time operating system, and the communication protocols include MQTT and ROS, for coordinating the operations of the traveling mechanism, the path navigation structure, and the weeding structure.

[0090] Among them, the traveling mechanism 1 includes four independently driven wheel assemblies, and each wheel assembly is equipped with a DC brushless motor (model: BLDC-200W), with a rated power of 200W, a speed range of 0-300 rpm, and an IP65 waterproof and dustproof rating. The motor controls the speed through a PWM signal (frequency 20 kHz, duty cycle 0% - 100%), and the adjustable range of the traveling speed is 0-1 m / s.

[0091] The path navigation structure 2 includes:

[0092] Searchlight: LED light source, illumination range 500-1000 lux, power 10W, adjustable angle ±30°;

[0093] LiDAR: model Lidar-X4, scanning frequency 10 Hz, ranging range 0.2-40 m, angular resolution 0.25°, generating a three-dimensional point cloud map, the formula is:

[0094] Px,y,z = {x i ,y i ,z i ∣x i = r i ·cosθ i ·cosφ i ,y i = r i ·sinθ i ·

[0095] cosφ i ,z i = r i ·sinφ i};

[0096] Among them, r i (m), θ i (rad), φ i (rad) are the distance, horizontal and vertical angles respectively, the point cloud density is 1000 points / m2, and the accuracy is ±0.05 m;

[0097] Laser obstacle avoidance radar: model SF-10, detection range 0.1-10 m, response time <50 ms, distance threshold D th = 0.5 m, triggering the A * algorithm path planning, the cost function is:

[0098]

[0099] Among them, (g(n)) and (h(n)) are in units of m;

[0100] Scene construction camera: Model Cam-20MP, 20 million pixels, field of view 120°, frame rate 30fps, combined with lidar to achieve SLAMo

[0101] The weeding structure 3 includes: Visual detection module: Binocular intelligent camera (Model: StereoCam-1080P), baseline length 120mm, focal length 8mm, resolution 1920×1080, frame rate 30fps, coordinate calculation formula:

[0102]

[0103] Among them, b = 120mm, f = 8mm, (d)(px) is the parallax, ((u,v))(px) is the pixel coordinate, accuracy ±2mm; Data processing module: GPU (Model: NVIDIAJetsonAGX), computing power 12TFLOPS, supporting two algorithms:

[0104] CNN+RNN+LSTM: Based on the CWD30 dataset, CNN (5 convolutional layers, 3×3 kernel, stride 1), feature extraction:

[0105] F l = ReLUW l *I + b l

[0106] RNN and LSTM (128 neurons), accuracy 95%;

[0107] YOLOv10+Mamba: Darknet-53 backbone, S = 13, B = 3, loss function:

[0108]

[0109] λ coord = 5, λ class = 1, mAP0.92;

[0110] The actuator includes: Three-axis robotic arm 31: Model ARM-3X, working radius 500mm, rotation range ±180°, accuracy ±0.1mm, inverse kinematics:

[0111] θ1 = arctan2y,x

[0112]

[0113] L1 = 200mm, L2 = 200mm, L3 = 100mm, h = 50mm;

[0114] Pump source laser 33: Model Laser-50W, wavelength 1064nm, power regulation: η = 0.85, Pin = 60W, A lens = 20mm 2 , A spot = 0.2mm 2 , d0 = 10mm;

[0115] High-magnification beam lens 32: Magnification is 10 times, and the spot diameter is 0.5mm.

[0116] [[ID=-- --]]Software control and communication module: Data transmission delay:

[0117] S data = 256KB, B = 100Mbps, N ops = 10 9 , F GPU =

[0118] 12 TFLOPS, T delay < 0.05s.

[0119] Error calibration: where R = 4mm 2 , with an error of ±1mm.

[0120] The following will be specifically described in conjunction with several embodiments;

[0121] Embodiment 1: Weeding operation in flat farmland;

[0122] Scenario: A flat farmland with an area of 100 m2, weeds are dandelions (height 50mm, root depth 15mm) and foxtails (height 40mm, root depth 10mm), soil humidity is 20%, and the operation time is during the day on March 11, 2025.

[0123] Step 1: Initialization;

[0124] Start the robot, set the illuminance of the searchlight to 500 lux (no strong light is needed during the day), the lidar scans at 10 Hz, collect point cloud data: r1 = 2.5m, θ1 = 0.5 rad, φ1 = 0.1 rad, calculate x1, y1, z1 = 2.38m, 1.19m, 0.25m, and generate a 100×100 point cloud grid. The scene construction camera synchronously collects images, the SLAM positioning accuracy is ±0.05mm, and the motor speed of the walking mechanism 1 is set to 150 rpm, with a speed of 0.5 m per second.

[0125] Step 2: Path planning;

[0126] The laser obstacle avoidance radar detects a stump ahead, the distance D = 0.3m < D th= 0.5m, trigger the A* algorithm. Starting coordinates ((0,0)), target (1)(10m,10m)), obstacle coordinates ((0.3m,0.2m)), calculate gn = 0.3m, hn = 14.1m, fn = 14.4m, the new path is the broken line 0,0 → 0.6m,0.4m → 10m,10m, and the time taken is 0.8s.

[0127] Step 3: Weed identification;

[0128] The binocular intelligent camera 34 captures an image and detects dandelions. The coordinates in the left image are u L , v L = 960px, 540px, and the coordinates in the right image are u R , v R = 940px, 540px, the parallax d = 20px, calculate X, Y, Z = 48mm, 24mm, 48mm. Identified by YOLOv10, the confidence level is 0.95, and the loss value Loss = 0.12.

[0129] Step 4: Weeding operation;

[0130] The three-axis robotic arm 31 receives the coordinates and calculates θ1 = arctan224,48 = 0.46rad, θ2 = 0.52rad, θ3 = 0.31rad, and the movement time is 0.15s.

[0131] The pump source laser 33 inputs d = 15mm, Burn for 0.5s, and the root removal rate is 100.

[0132] Step 5: Calibration and communication;

[0133] Calibrate using Kalman filter, the observed value z k = 48.5mm, the predicted value Gain K k = 0.6, after calibration The error is ±1mm.

[0134] Transmit 256KB of data using MQTT, with a delay of 0.02s.

[0135] Result: The operation takes 15 minutes, the weeding rate is 98%, and the coverage rate is 100%.

[0136] Example 2: Weeding operation in complex terrain;

[0137] A hilly farmland with an area of 50㎡, the weeds are amaranth (height 60mm, root depth 20mm) and goosegrass (height 30mm, root depth 12mm), the slope is 15°, the height difference is 50mm, and the operation time is at night on March 11, 2025.

[0138] Step 1: Initialization;

[0139] Adjust the searchlight illuminance to 1000 lux, scan with the lidar, point cloud data: r1 = 1.8 m, θ1 = 0.8 rad, φ1 = 0.2 rad, calculate x1, y1, z1 = 1.62 m, 1.32 m, 0.36 m, and generate a 50×50 grid.

[0140] Camera-assisted SLAM, positioning error ±0.06 m. Motor speed 200 rpm, speed 0.4 m / s.

[0141] Step 2: Path planning;

[0142] The obstacle avoidance radar detects a stone, D = 0.4 m < 0.5 m, A * Algorithm planning, starting point ((0,0)), target ((5 m, 5 m)), obstacle (1), (0.4 m, 0.3 m)), gn = 0.4 m, hn = 7.0 m, fn = 7.4 m, new path 0,0 → 0.8 m, 0.6 m → 5 m, 5 m, time-consuming 0.9 S.

[0143] Step 3: Weed identification;

[0144] The binocular intelligent camera 34 detects goosegrass, then there are:

[0145] (u L , v L ) = 950 px, 530 px, u R , v R = 932 px, 530 px, d = 18 px, calculate X, Y, Z = 53.3 mm, 26.7 mm, 53.3 mm.

[0146] CNN + RNN + LSTM identification, feature map F5 = 0.92, accuracy 96%.

[0147] Step 4: Weeding operation;

[0148] The calculation of the three-axis robotic arm 31 is θ1 = 0.46 rad, θ2 = 0.60 rad, θ3 = 0.28 rad, and the movement time is 0.18 s.

[0149] The laser d = 12 mm, Burn for 0.6 s, removal rate 99%.

[0150] Step 5: Calibration and communication;

[0151] Kalman filter, z k = 53.5 mm, K k = 0.5, Error: ±1mm. ROS transmits data with a delay of 0.025s.

[0152] Result: It takes 12 minutes, with a weeding rate of 95%, and it adapts well to slope changes.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent weeding robot, characterized in that, Comprising: A traveling mechanism (1) for driving the entire robot to travel. The traveling mechanism includes four independently driven wheel assemblies, and each wheel assembly is equipped with a DC brushless motor; A path navigation structure (2) for performing path navigation and sending instructions to the traveling mechanism to control its travel. The path navigation structure includes: a searchlight, a lidar, and a laser obstacle avoidance radar. Among them, the searchlight and the lidar are both arranged on the top of the housing, and the laser obstacle avoidance radar is arranged on the front end face of the vehicle frame. The housing is connected to the vehicle frame by bolts; A weeding structure (3) for removing weeds on the ground. The weeding structure includes: A vision detection module, using a binocular intelligent camera, for obtaining the spatial position information of weeds; A data processing module, based on a high-performance GPU with a computing power of 12 TFLOPS, running a weed recognition algorithm and performing coordinate calculation and error calibration; An execution mechanism, including a pump source laser (33), a high-power beam lens (32), and a three-axis robotic arm (31); the binocular intelligent camera (34), the pump source laser (33), and the high-power beam lens (32) are all fixed on the output shaft of the three-axis robotic arm (31), and the three-axis robotic arm is fixedly installed on the vehicle frame; A software control and communication module, receiving vision signals based on a high-performance GPU, running a real-time operating system, and the communication protocols include MQTT and ROS, for coordinating the operation of the traveling mechanism, the path navigation structure, and the weeding structure.

2. The intelligent weeding robot according to claim 1, characterized in that The path navigation structure constructs a three-dimensional point cloud map of the environment through the lidar, and performs SLAM in combination with the scene construction camera. The point cloud generation formula is: Px,y,z = {x i , y i , z i | x i = r i · cosθ i · cosφ i , y i = r i · sinθ i · cosφ i ,z i =r i ·sinφ i}; where r i is the distance measured by the lidar; θ i is the horizontal scanning angle in radians; φ i is the vertical scanning angle in radians; x i , y i , z i are the point cloud coordinates in meters; i is the sampling point number; the point cloud density is 1000 points per square meter.

3. The intelligent weeding robot according to claim 1, characterized in that, The vision detection module calculates the spatial coordinates of weeds through a binocular intelligent camera. The coordinate calculation formula is: Wherein, b is the baseline length, with a value of 120 mm and the unit of millimeter; f is the camera focal length, with a value of 8 mm and the unit of millimeter; d is the parallax of the weed feature points in the left and right images, with the unit of pixel; u, v are the pixel coordinates on the image plane, with the unit of pixel; X, Y, Z are the spatial coordinates of the weeds, with the unit of millimeter; the calculation accuracy of the binocular intelligent camera is ±2 mm, and the parallax d is obtained through a feature point matching algorithm.

4. The intelligent weeding robot according to claim 1, characterized in that, The data processing module uses a CNN based on the CWD30 dataset to fuse RNN and LSTM algorithms for weed recognition. Among them, the CNN contains 5 convolutional layers and 3 pooling layers, the convolutional kernel size is 3×3 pixels, the stride is 1, and the feature extraction formula is: F l = ReLUW l * I + b l ; Among them, F l is the feature map of the (I)-th layer, with the unit of dimensionless; ReLU is the activation function, and: ReLUx = max(0, x); W l is the convolutional kernel weight matrix, with the unit of dimensionless; I is the input image, with the unit of pixel value (0 - 255); b l is the bias, with the unit of dimensionless; * is the convolution operation; RNN and LSTM process time series data, the number of neurons in the hidden layer is 128, and the recognition accuracy is 95%.

5. The intelligent weeding robot according to claim 1, wherein, The data processing module uses a weed recognition system based on YOLOv10 and Mamba. Among them, the YOLOv10 network contains a Darknet-53 backbone network, the number of detection boxes is 3, and the loss function is: Among them, λ coord is the coordinate loss weight, with a value of 5 and dimensionless; λ class : classification loss weight, value is 1, dimensionless; S is the grid size, with a value of 13 and dimensionless; B is the number of bounding boxes, with a value of 3 and dimensionless; The target exists indicator, which takes a value of 0 or 1; x i ,y i and are the predicted and true coordinates in pixels; p i c and are class probabilities, ranging from 0 to 1; The average precision mAP of the optimized Mamba model is greater than 0.

92.

6. The intelligent weeding robot according to claim 1, wherein The three-axis robotic arm controls the motion trajectory through inverse kinematics. The joint angle calculation formula is: Among them, X, Y, and Z are the target positions of the weeds, with the unit of millimeters; h is the height of the base, with a value of 50 mm and the unit of millimeters; L1, L2, and L3 are the lengths of each segment of the robotic arm, which are 200 mm, 200 mm, and 100 mm respectively, with the unit of millimeters; θ1, θ2, and θ3 are the joint angles, with the unit of radians; the response time is less than 0.2 s.

7. The intelligent weeding robot according to claim 1, characterized in that, The output power of the pump source laser is dynamically adjusted according to the weed depth, and the calculation formula is: Among them, P out is the output power in watts; η is the conversion efficiency, with a value of 0.85 and dimensionless; P in is the input power, with a value of 60 W and the unit of watt; A lens is the effective area of the lens, with a value of 20 mm 2 , and the unit is square millimeters; A spot is the area of the focused spot, with a value of 0.2 and the unit of square millimeter; d is the depth of the weed root, with the unit of millimeters; d0 is the reference depth, with a value of 10 mm and the unit of millimeters; the power adjustment range is 20 - 50 W.

8. The intelligent weeding robot according to claim 1, characterized in that The software control and communication module realizes data transmission through the MQTT protocol, and the message packet size is 256 KB. The transmission delay calculation formula is: T delay is the total delay in seconds, and it is required to be less than 0.05 s; S data is the data volume, with a value of 256 KB and the unit of bits; B is the bandwidth, with a value of 100 Mbps and the unit of bits per second; N ops is the amount of computation, with a value of 10 9 times of computation, with the unit of times; F GPU The computing power is 12 TFLOPS, measured in times per second. The calculation result is approximately 0.02s, which meets the latency requirement.

9. The intelligent weeding robot according to claim 1, characterized in that The distance threshold D for obstacle detection by the lidar for obstacle avoidance in the path navigation structure th = 0.5 m. When the detected distance D < D th , trigger the A * algorithm to re-plan the path. The path cost function is as follows: fn = gn + hn; Among them, f(n) is the total cost, with the unit of meters; (g(n)) is the actual distance from the starting point to the current node, with the unit of meters; (h(n)) is the Euclidean distance from the current node to the target, with the unit of meters, and the calculation formula is: Among them, x n , y n are the coordinates of the current node, with the unit of meter; x g , y g are the target coordinates, with the unit of meter; the planning time is less than 1 s, and the resolution is 0.1 m.

10. The intelligent weeding robot according to claim 1, characterized in that, The weeding structure calibrates the coordinate error through the Kalman filtering algorithm when performing tasks, and the state update formula is: in, is the state vector after calibration, in millimeters; is the predicted state vector, in millimeters; K k is the Kalman gain matrix, dimensionless; z k is the observation value, in millimeters; H is the observation matrix, dimensionless; P k|k-1 is the prediction covariance matrix, in mm 2 ; R is the measurement noise covariance, unit is mm 2 , typical value is 4mm 2 ;The error is reduced to ±1mm.

Images