Adaptive fruit harvesting control methods, systems, media, and robots
By using a hybrid sphere fitting and vector-guided control method, the problems of inaccurate fruit positioning and suboptimal motion planning in tomato harvesting robots were solved, achieving a high-precision and safe fruit grasping and harvesting process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies in tomato harvesting robots suffer from problems such as inaccurate fruit positioning, poor scale adaptability, and suboptimal motion planning, leading to issues like slippage during grasping, crushing damage, and collisions with the robotic arm.
A hybrid sphere fitting and vector-guided control method is adopted. The depth point cloud of the fruit and stalk is obtained through instance segmentation. An adaptive dynamic search domain is constructed. Combined with the stalk growth vector and the inverse kinematics solution of the robotic arm, high-precision fruit grasping is achieved.
It improves the precision and safety of fruit grasping, avoids robotic arm collisions and system shutdowns, and achieves an efficient and safe harvesting process.
Smart Images
Figure CN122480937A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural automation and robotics, specifically to an adaptive fruit harvesting control method based on hybrid sphere fitting and vector guidance, and more particularly to an adaptive tomato harvesting control method that combines computer vision perception, 3D point cloud processing, and robotic arm motion planning. Background Technology
[0002] With the rapid development of facility agriculture, fruit-harvesting robots have become a key piece of equipment for solving labor shortages. However, in actual unstructured greenhouse environments, harvesting robots face numerous technical challenges.
[0003] Tomatoes, a widely cultivated fruit and vegetable crop in greenhouse agriculture, have attracted considerable attention in recent years for the research of harvesting robots. Tomato fruits are typically spherical, varying in size, and grow densely in clusters, with the stem located at the top of the fruit, exhibiting diverse postures. Existing technologies typically utilize two-dimensional object detection algorithms (such as YOLO) to obtain the rectangular bounding box of the tomato, using its center as the grasping target. However, the varying shapes, sizes, and growth postures of tomatoes mean that relying solely on the surface center point for positioning often results in the robotic arm's end effector failing to completely cover the fruit, leading to slippage or crushing damage. Furthermore, in the fruit-stem matching stage, existing methods often employ fixed pixel distance thresholds, which cannot adapt to the scaling issues caused by changes in the distance between the camera and the target, easily resulting in missed detections or false matches.
[0004] In terms of motion planning, traditional picking strategies often lack optimization of the operation sequence, and random grasping can easily lead to collisions between the robotic arm and unpicked fruit in front. At the same time, a single strategy of aligning the fruit stem for grasping may result in unsolvable or singular inverse kinematics of the robotic arm at certain extreme angles, causing the system to stop.
[0005] Therefore, there is an urgent need for a control method that can accurately recover the three-dimensional geometric appearance of a fruit, has scale adaptability, and is robust to motion planning. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a fruit adaptive harvesting control method based on hybrid sphere fitting and vector guidance.
[0007] An adaptive fruit harvesting control method according to the present invention includes the following steps: Step S1: Obtain an RGB image containing the fruit and stem and the corresponding depth map, and use a pre-trained image segmentation model to perform instance segmentation on the RGB image to obtain the segmentation results and confidence scores of the fruit and stem. Step S2: Based on the confidence and depth information, the segmented fruits are filtered and sorted to generate a job time sequence queue; Step S3: For the current fruit in the job time queue, combine its segmentation results and depth map, and use a hybrid sphere fitting algorithm to calculate the fruit's grasping center. Step S4: Construct an adaptive dynamic search domain based on the visual radius of the current fruit, perform fruit stem matching within the dynamic search domain, establish the association between the fruit and the fruit stem, and calculate the fruit stem growth vector; Step S5: Based on the validity of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, an adaptive hierarchical decision is made between the vector guidance strategy and the position priority strategy to control the end effector of the robotic arm to complete the picking action.
[0008] Preferably, step S2 specifically includes: Set a confidence threshold and remove detection targets with a confidence level below that threshold; Extract the depth values of the remaining valid target center points, and generate an initial job queue in ascending order of depth values; For multi-target scenarios with overlap or interference, mask overlap verification or 3D point cloud interference verification is introduced, and the priority of the initial job queue is adjusted so that non-interfering targets are moved to the first position in the queue.
[0009] Preferably, step S3 specifically includes: Extract the corresponding depth point cloud set from the depth map using the binary mask of the current fruit; The point cloud set was fitted with a three-dimensional sphere using the least squares method to calculate the geometric center coordinates and physical radius of the sphere. Determine whether the fitted radius is within the preset physical constraint range. If so, use the geometric center of the sphere as the grasping center; otherwise, trigger the backtracking mechanism and use the coordinates of the center point of the fruit surface after compensating for a fixed radius along the camera optical axis as the grasping center.
[0010] Preferably, step S4 specifically includes: Calculate the visual radius of the current fruit detection box; A dynamic circular search domain is constructed based on the visual radius and preset search coefficients; Within the dynamic circular search domain, candidate fruit stalks are screened, and the fruit stalks closest to the center of the fruit are selected to establish an association pair. By using camera intrinsic parameters to backproject the associated fruit stalk center into three-dimensional space, a unit growth vector pointing from the fruit grasping center to the fruit stalk center is constructed as the fruit stalk growth vector.
[0011] Preferably, step S5 specifically includes: If the fruit stalk growth vector is valid and the inverse kinematics of the robotic arm has a solution, then the vector guidance strategy is executed: construct the end-target rotation matrix so that the end-target approach axis of the robotic arm is collinear with the fruit stalk growth vector, thereby achieving a follow-along envelope grasping; If the fruit stalk growth vector is invalid or the inverse kinematics has no solution, then the position priority strategy is executed: maintain the current posture of the robotic arm end effector, only adjust the end effector position coordinates to the grasping center, and complete the downgraded grasping.
[0012] Preferably, the image segmentation model is a semantic segmentation model based on YOLOv11-seg, and the segmentation result includes contour masks of the fruit and its stem.
[0013] Preferably, in step S5, after the grasping is completed, the step of twisting and separating the fruit stem is further included: keeping the position of the end of the robotic arm unchanged, controlling the end joint to rotate at a preset angle, and using shearing force to separate the fruit stem.
[0014] An adaptive fruit harvesting control system according to the present invention includes: An RGB-D camera is used to acquire RGB images containing the fruit and its stem, as well as corresponding depth maps. The image segmentation module is used to perform instance segmentation on the RGB image using a pre-trained image segmentation model to obtain the segmentation results and confidence scores of the fruit and fruit stalk. The job sorting module is used to filter and sort the segmented fruits based on the confidence and depth information, and generate a job time sequence queue. The positioning and fitting module is used to calculate the grasping center of the current fruit in the job time queue by combining its segmentation results and depth map. The fruit stalk matching module is used to construct an adaptive dynamic search domain based on the visual radius of the current fruit, perform fruit stalk matching within the dynamic search domain, establish the association between the fruit and the fruit stalk, and calculate the fruit stalk growth vector. The hierarchical decision module is used to make adaptive hierarchical decisions between the vector guidance strategy and the position priority strategy based on the effectiveness of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, and generate control commands. The robotic arm and end effector are used to receive the control commands and perform the picking action.
[0015] According to the present invention, a computer-readable storage medium is provided thereon storing a computer program that, when executed by a processor, implements the steps of the adaptive fruit picking control method.
[0016] A fruit-harvesting robot according to the present invention includes: Multi-degree-of-freedom robotic arm; An end effector installed at the end of a robotic arm; RGB-D camera; A controller, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the adaptive fruit picking control method.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The hybrid sphere fitting algorithm proposed in this invention makes full use of instance segmentation mask to accurately extract the depth point cloud of the fruit region, and directly calculates the geometric center and physical radius of the fruit through least squares sphere fitting. Compared with the traditional positioning method based on the center of two-dimensional bounding box, it can more realistically reflect the complete geometric shape of the fruit.
[0018] 2. This invention constructs a dynamic search domain matching mechanism based on the visual radius of the fruit. By calculating the visual radius of the fruit detection box and multiplying it by a preset search coefficient, a dynamic search range independent of the camera distance is automatically formed, which solves the drawback of traditional fixed pixel thresholds that are prone to missed detection or mismatch when the shooting distance changes.
[0019] 3. This invention employs a depth-first sorting logic to generate a task sequence queue from nearest to farthest, allowing the robotic arm to prioritize picking the nearest target fruit. This implicitly implements a "from outside to inside, from near to far" de-occlusion strategy in terms of task order. Furthermore, for densely growing scenarios, mask overlap verification and 3D point cloud interference verification are introduced to dynamically adjust task priorities, effectively avoiding the risk of collisions between the robotic arm and unpicked fruit during operation, ensuring the safety and continuity of the picking process.
[0020] 4. When the fruit stalk growth vector is valid and the inverse kinematics has a solution, this invention employs a vector-guided strategy to achieve ideal grasping along the fruit stalk direction; when the fruit stalk is missing or the inverse kinematics has no solution, it automatically downgrades to a position-priority strategy, maintaining the current posture and only adjusting the position coordinates to complete a minimum grasp. This hierarchical mechanism ensures efficient harvesting under ideal posture while avoiding system downtime due to single-path failure. Attached Figure Description
[0021] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of a tomato dataset annotation provided by the present invention; Figure 3 This is a schematic diagram of the hybrid sphere fitting and positioning algorithm in this invention; Figure 4This is a schematic diagram illustrating the dynamic search domain matching principle of the two-dimensional image plane in this invention. Figure 5 This is a geometric diagram illustrating the growth vector of the fruit stalk in this invention. Detailed Implementation
[0022] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0023] This embodiment takes tomato harvesting as an example. Addressing the problems of large fruit positioning errors, difficult component matching, and limited robotic arm motion planning in unstructured environments, it provides an adaptive tomato harvesting control method based on hybrid sphere fitting and vector guidance. This method relies on a harvesting robot platform comprising an RGB-D camera, a multi-degree-of-freedom robotic arm, a three-finger flexible gripper at the end effector, and a main control computer. The overall process of this method is as follows: Figure 1 As shown, the specific execution steps are as follows: Step S1: Construct a tomato instance segmentation dataset, and use this dataset to train the image instance segmentation model to obtain a tomato image segmentation model, such as... Figure 2 As shown.
[0024] Step S1 is as follows: Step S1.1: Acquire target images. Collect RGB images of tomatoes under different growth stages, lighting conditions, and occlusion scenarios at an agricultural planting base. Use annotation tools to annotate the polygonal contours of ripe tomatoes, unripe tomatoes, and stems to construct a multi-class segmentation dataset. Expand the sample through methods such as random rotation and brightness adjustment to obtain a tomato image segmentation dataset, and divide it into training, validation, and test sets in a ratio of 7:2:1.
[0025] Step S1.2: Train the YOLOv11-seg semantic segmentation model using the tomato image segmentation dataset, enabling it to recognize and segment the contours of all tomato fruits and stems within the field of view, thus obtaining a trained tomato image segmentation model. The training platform uses a high-performance GPU server, and iterative training continues until the model converges.
[0026] Step S2: Obtain the RGB image of the tomato to be processed and the corresponding depth map. Input the RGB image into the tomato image segmentation model, and the model outputs the corresponding tomato segmentation result. Perform double filtering and depth sorting on the segmentation result to construct the anti-collision operation sequence.
[0027] Step S2 is as follows: Step S2.1: Set the confidence threshold Iterate through all "ripe-tomato" targets output by the model, and remove those with a confidence level less than 1. The test results are used to filter out immature or falsely detected targets.
[0028] Step S2.2: Extract the Z-axis depth value of the remaining valid target center points in the camera coordinate system. If the depth value exceeds the effective working range (less than 0.1m or greater than 1.0m), it will be rejected.
[0029] Step S2.3: Arrange the effective targets according to The targets are sorted in ascending order (from nearest to farthest). To address the issue of multi-target overlap in densely growing scenes, a mask overlap verification strategy is introduced: for the first target tomato in the queue, the intersection-union ratio (IUU) or contour approximation distance between its mask boundary and the masks of the other tomatoes is calculated. If the first target is severely occluded or there is 3D point cloud interference on its approach trajectory, a priority degradation strategy is triggered, and the next closest fruit without interference is extracted and placed at the first position for execution. This supplementary strategy effectively avoids collisions or deadlocks caused by dense overlapping.
[0030] Step S3: For the current target tomato in the job queue, combining the segmentation mask and depth map, a hybrid sphere fitting algorithm is used to calculate the geometric center and physical radius of the fruit, such as... Figure 3 As shown.
[0031] Step S3 is as follows: Step S3.1: Using the binary mask of the target tomato as an index, extract a set of depth point clouds containing only the fruit region from the depth map. After extraction, a validity check is performed. If the number of valid point clouds is... Less than the set point cloud number threshold If the depth is severely lacking, the spherical fitting is skipped, triggering the rollback mechanism in step 3.3.
[0032] Step S3.2: Apply the least squares method to the point cloud set Perform a three-dimensional spherical fitting. Establish the standard spherical equation. To improve solution efficiency, the nonlinear equation is algebraically expanded and transformed into... The linear overdetermined system of equations. Through matrix... Perform singular value decomposition or directly solve for the parameter vector using the least squares method. Thus, the geometric center of the fruit can be directly determined. and physical radius .
[0033] Step S3.3: Perform robustness verification. Set a reasonable physical radius range. .like If the fit is successful, then the fitting is considered successful. Used as the center of capture; if the conditions are not met (such as depth loss due to lighting, or fitting divergence caused by sparse point cloud), a backoff mechanism is triggered: read the coordinates of the center point of the target surface. Compensation of fixed empirical radius along the camera optical axis ,Right now .
[0034] Step S4: Construct an adaptive dynamic search domain based on the visual radius, perform fruit stalk matching within the dynamic domain, establish the association between the fruit and the stalk, and calculate the stalk growth vector, such as... Figure 4 As shown.
[0035] Step S4 is as follows: Step S4.1: Calculate the width of the target tomato detection box. and high Determine the visual radius :
[0036] Step S4.2: Set the dynamic search threshold :
[0037] Among them, search coefficient These are adjustable parameters set based on the morphological characteristics of tomato varieties. The tomato fruit stalk is usually located at the top of the fruit and close to the center of mass; in this embodiment, a coefficient is used. The effective search domain is defined as the area extending outward from the center of the fruit to twice the visual radius. This threshold utilizes the visual radius to achieve dynamic scaling and possesses distance-scale invariance; when changing cameras or targeting varieties with long fruit stalks (such as clustered cherry tomatoes), it only needs to be adjusted proportionally based on the statistical average proportion of crop morphology. The value is sufficient.
[0038] Step S4.3: Traverse all "stem" category targets in the field of view and calculate the pixel Euclidean distance between the center of each target and the center of the target tomato. Filter to meet Candidate fruit stalks were selected. The smallest one serves as the associated peduncle.
[0039] Step S4.4: Use camera intrinsic parameters to backproject the center of the associated fruit stalk into three-dimensional coordinates. Construct a unit growth vector pointing from the fruit grasping center to the fruit stalk center. ,like Figure 5 As shown.
[0040]
[0041] Step S5: Based on the validity of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, an adaptive hierarchical decision is made between the vector guidance strategy and the position priority strategy to control the end effector of the robotic arm to complete the picking action.
[0042] Step S5 is as follows: Step S5.1: Strategy S1 (Vector-oriented): If a fruit stalk is successfully matched in step 4 and Construct the end target rotation matrix This makes the approach axis (Z-axis) of the robotic arm's end-effector coordinate system align with... Collinear. Call the inverse kinematics solver to verify that the pose has a solution and is free of singularities. If the verification passes, the robotic arm moves along... The direction naturally envelops the fruit.
[0043] Step S5.2: Strategy S2 (Position Priority): If no fruit stalk is matched or Strategy S1 results in no solution for inverse kinematics, the system automatically downgrades to Strategy S2. Maintaining the current posture of the robotic arm's end effector (usually horizontal or vertical), only the end effector's position coordinates are adjusted to the grasping center determined in Step 3. .
[0044] Step S5.3: Control the robotic arm to execute the planned path, and close the three-finger gripper to cover the fruit. Then, keeping the end position unchanged, control the end joint to rotate 90° to 180°, using shearing force to twist and separate the fruit stem, completing the harvest. Step S6: Repeat steps S3 to S5 until all targets in the job queue have been processed.
[0045] The fruit adaptive harvesting control method based on hybrid sphere fitting and vector guidance proposed in this invention is not only applicable to tomato harvesting scenarios, but can also be extended to other fruit and vegetable crops with spherical or near-spherical shapes, such as peppers, cucumbers, apples, and citrus.
[0046] The present invention also provides a fruit adaptive harvesting control system, which can be implemented by executing the process steps of the fruit adaptive harvesting control method. That is, those skilled in the art can understand the fruit adaptive harvesting control method as a preferred embodiment of the fruit adaptive harvesting control system.
[0047] The present invention provides a fruit adaptive harvesting control system, comprising: An RGB-D camera is used to acquire RGB images containing the fruit and its stem, as well as corresponding depth maps. The image segmentation module is used to perform instance segmentation on the RGB image using a pre-trained image segmentation model to obtain the segmentation results and confidence scores of the fruit and fruit stalk. The job sorting module is used to filter and sort the segmented fruits based on the confidence and depth information, and generate a job time sequence queue. The positioning and fitting module is used to calculate the grasping center of the current fruit in the job time queue by combining its segmentation results and depth map. The fruit stalk matching module is used to construct an adaptive dynamic search domain based on the visual radius of the current fruit, perform fruit stalk matching within the dynamic search domain, establish the association between the fruit and the fruit stalk, and calculate the fruit stalk growth vector. The hierarchical decision module is used to make adaptive hierarchical decisions between the vector guidance strategy and the position priority strategy based on the effectiveness of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, and generate control commands. The robotic arm and end effector are used to receive the control commands and perform the picking action.
[0048] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of adaptive fruit harvesting control.
[0049] The present invention also provides a fruit-harvesting robot, comprising: Multi-degree-of-freedom robotic arm; An end effector installed at the end of a robotic arm; RGB-D camera; A controller, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the adaptive fruit picking control method.
[0050] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.
[0051] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A fruit adaptive picking control method characterized by, Includes the following steps: Step S1: Obtain an RGB image containing the fruit and stem and the corresponding depth map, and use a pre-trained image segmentation model to perform instance segmentation on the RGB image to obtain the segmentation results and confidence scores of the fruit and stem. Step S2: Based on the confidence and depth information, the segmented fruits are filtered and sorted to generate a job time sequence queue; Step S3: For the current fruit in the job time queue, combine its segmentation results and depth map, and use a hybrid sphere fitting algorithm to calculate the fruit's grasping center. Step S4: Construct an adaptive dynamic search domain based on the visual radius of the current fruit, perform fruit stem matching within the dynamic search domain, establish the association between the fruit and the fruit stem, and calculate the fruit stem growth vector; Step S5: Based on the validity of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, an adaptive hierarchical decision is made between the vector guidance strategy and the position priority strategy to control the end effector of the robotic arm to complete the picking action.
2. The fruit adaptive picking control method according to claim 1, wherein, Step S2 specifically includes: Set a confidence threshold and remove detection targets with a confidence level below that threshold; Extract the depth values of the remaining valid target center points, and generate an initial job queue in ascending order of depth values; For multi-target scenarios with overlap or interference, mask overlap verification or 3D point cloud interference verification is introduced, and the priority of the initial job queue is adjusted so that non-interfering targets are moved to the first position in the queue.
3. The method of claim 1, wherein, Step S3 specifically includes: Extract the corresponding depth point cloud set from the depth map using the binary mask of the current fruit; The point cloud set is fitted with a three-dimensional sphere using the least squares method to calculate the geometric center coordinates and physical radius of the sphere. Determine whether the fitted radius is within the preset physical constraint range. If so, use the geometric center of the sphere as the grasping center; otherwise, trigger the backtracking mechanism and use the coordinates of the center point of the fruit surface after compensating for a fixed radius along the camera optical axis as the grasping center.
4. The method of claim 1, wherein, Step S4 specifically includes: Calculate the visual radius of the current fruit detection box; A dynamic circular search domain is constructed based on the visual radius and preset search coefficients; Within the dynamic circular search domain, candidate fruit stalks are screened, and the fruit stalks closest to the center of the fruit are selected to establish an association pair. By using camera intrinsic parameters to backproject the center of the associated fruit stalk into three-dimensional space, a unit growth vector pointing from the fruit grasping center to the center of the fruit stalk is constructed as the fruit stalk growth vector.
5. The method of claim 1, wherein, Step S5 specifically includes: If the fruit stalk growth vector is valid and the inverse kinematics of the robotic arm has a solution, then the vector guidance strategy is executed: construct the end-target rotation matrix so that the end-target approach axis of the robotic arm is collinear with the fruit stalk growth vector, thereby achieving a follow-along envelope grasping; If the fruit stalk growth vector is invalid or the inverse kinematics has no solution, then the position priority strategy is executed: maintain the current posture of the robotic arm end effector, only adjust the end effector position coordinates to the grasping center, and complete the downgraded grasping.
6. The fruit adaptive harvesting control method according to claim 1, characterized in that, The image segmentation model is a semantic segmentation model based on YOLOv11-seg, and the segmentation results include contour masks of the fruit and its stem.
7. The fruit adaptive harvesting control method according to claim 1, characterized in that, In step S5, after the grasping is completed, the step of twisting and separating the fruit stem is also included: keeping the position of the end of the robotic arm unchanged, controlling the end joint to rotate at a preset angle, and using shearing force to separate the fruit stem.
8. A fruit adaptive harvesting control system, characterized in that, include: An RGB-D camera is used to acquire RGB images containing the fruit and its stem, as well as corresponding depth maps. The image segmentation module is used to perform instance segmentation on the RGB image using a pre-trained image segmentation model to obtain the segmentation results and confidence scores of the fruit and fruit stalk. The job sorting module is used to filter and sort the segmented fruits based on the confidence and depth information, and generate a job time sequence queue. The positioning and fitting module is used to calculate the grasping center of the current fruit in the job time queue by combining its segmentation results and depth map. The fruit stalk matching module is used to construct an adaptive dynamic search domain based on the visual radius of the current fruit, perform fruit stalk matching within the dynamic search domain, establish the association between the fruit and the fruit stalk, and calculate the fruit stalk growth vector. The hierarchical decision module is used to make adaptive hierarchical decisions between the vector guidance strategy and the position priority strategy based on the effectiveness of the fruit stalk growth vector and the inverse kinematics solution state of the robotic arm, and generate control commands. The robotic arm and end effector are used to receive the control commands and perform the picking action.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 7.
10. A fruit-harvesting robot, characterized in that, include: Multi-degree-of-freedom robotic arm; An end effector installed at the end of a robotic arm; RGB-D camera; A controller, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the method according to any one of claims 1 to 7.