Three-dimensional reconstruction based on binocular vision and obstacle avoidance path planning method for weeding robot arm

Abstract

The application relates to the technical field of agricultural robot control, and discloses a three-dimensional reconstruction and obstacle avoidance path planning method of a weeding mechanical arm based on binocular vision, which comprises the following steps: monitoring the mean value of an optical flow, triggering the mechanical arm to generate active excitation vibration under a steady state; eliminating transmission error to extract an environment net response optical flow, combining a point cloud to construct a three-dimensional stiffness map distinguishing rigid and flexible; constructing an anisotropic cost field according to the main direction of the texture of a flexible area, generating an optimal penetration path conforming to the texture; fusing a visual divergence field and a momentum observer contact torque when the path is executed, outputting real-time interaction features; judging effective flexible penetration, rigid blocking or false contact states according to the real-time interaction features, and adaptively adjusting impedance control parameters. The application decouples the stiffness of the environment through active perception, utilizes the texture to guide flexible penetration, and uses a visual contact fusion mechanism, so that the problems of low working efficiency and easy damage to crops of the mechanical arm in a dense vegetation environment are solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural robot control technology, specifically to a method for three-dimensional reconstruction and obstacle avoidance path planning for a weeding robotic arm based on binocular vision. Background Technology

[0002] With the development of precision agriculture, robotic arms for weeding are increasingly being used in field operations, replacing traditional manual weeding or chemical herbicide spraying. In unstructured agricultural environments, robotic arms need to locate and perform tasks amidst dense crops and weeds. However, the complex field environment, with crops and weeds intertwined and obstructing each other, places demands on the robot's environmental perception and motion planning capabilities.

[0003] Existing environmental perception systems for agricultural robots mostly employ passive binocular vision or LiDAR for 3D scene reconstruction. While these methods can acquire geometric depth information of the environmental surface, they cannot perceive the physical stiffness properties of objects. In the data collected by visual sensors, rigid crop stalks, support rods, and flexible crop leaves all appear as point cloud data occupying space. Due to the difficulty in distinguishing the stiffness properties of obstacles, existing perception systems typically treat all geometric protrusions within the field of view as rigid obstacles. This approach results in a lack of feasible space for constructing environmental maps in densely vegetated areas, limiting the robotic arm's operating range.

[0004] In path planning, traditional obstacle avoidance algorithms primarily follow the collision-free principle, searching for geometric paths that avoid all obstacles. However, when crop canopy closure is high, weeds are often obscured or surrounded by crop leaves. If the collision-free principle is strictly followed, the robotic arm may struggle to find a feasible path to the target point, or the resulting path may involve excessively long detours, reducing operational efficiency. Furthermore, existing planning algorithms typically do not consider the biomorphological characteristics and growth direction of plants, making it easy for the robotic arm to vertically cut or reverse-compress plant fibers during movement, causing physical damage to the crop.

[0005] In terms of motion control and interactive safety, when the robotic arm penetrates deep into vegetation, the camera mounted at its end is easily obstructed by branches and leaves, leading to visual feedback failure or increased data noise. Relying solely on force sensors for collision detection suffers from response lag and an inability to predict contact trends. Existing control strategies struggle to effectively integrate visual and force information in complex contact operations, failing to distinguish in real time whether the robotic arm is pushing aside flexible leaves or encountering rigid branches. This makes it difficult for the control system to dynamically adjust the robotic arm's compliance characteristics to balance operational efficiency and safety. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for three-dimensional reconstruction and obstacle avoidance path planning for a weeding robotic arm based on binocular vision. This method solves the problems of existing agricultural robot environmental perception systems being unable to distinguish between the rigid and flexible properties of crops, and obstacle avoidance strategies treating all obstacles as rigid bodies. As a result, the robotic arm cannot effectively utilize the gaps between leaves, leading to low work efficiency and easy damage to crops when working in dense vegetation.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for obstacle avoidance path planning for a weeding robotic arm based on binocular vision-based 3D reconstruction includes the following steps: The industrial control computer monitors the global optical flow data collected by the binocular depth camera. When it determines that the average global optical flow is lower than the preset static judgment threshold, it injects a sinusoidal position disturbance command into the six-degree-of-freedom robotic arm mounted on the mobile chassis, which drives the end effector to generate an active excitation vibration signal. The stiffness mapping module receives the active excitation vibration signal and the synchronized joint encoder data, eliminates the transmission error of the robotic arm and extracts the net response optical flow of the environment, and combines the point cloud geometric information to generate a three-dimensional stiffness map containing the distribution of rigid stems and flexible leaves. The path planning module constructs a directional cost field based on the main texture direction of the flexible region in the three-dimensional stiffness map, and searches and generates an optimal penetration path in three-dimensional space based on the principle of energy optimization, which is parallel to the texture direction of the plant leaf. The six-degree-of-freedom robotic arm executes the optimal penetration path, and the interactive control module calculates the visual optical flow divergence field of the contact area and the external contact torque based on the momentum observer in real time, and outputs real-time interactive features characterizing the current contact state. The industrial control computer determines the current state as an effective flexible penetration state, a rigid blocking state, or a false contact state based on the real-time interaction characteristics. Accordingly, it adjusts the impedance control parameters of the six-degree-of-freedom robotic arm, triggers emergency stop replanning, or maintains the current motion state, and generates the final control execution command.

[0008] Preferably, during the monitoring of global optical flow data and the generation of active excitation vibration signals, the industrial control computer processes the collected global optical flow data using the global optical flow mean calculation formula to obtain the global optical flow mean reflecting the overall intensity of environmental motion. Only when the environment is in a steady state is a sinusoidal position perturbation command in Cartesian space generated using a sinusoidal position perturbation formula including a smooth envelope function and a preset crop resonance frequency. This process enhances the deformation response of the plant through frequency matching and avoids the impact of step signals on the joints of the robotic arm by using the envelope function. The static determination threshold is 0.3 to 0.8 pixels / frame.

[0009] Preferably, in the process of processing kinematic data to eliminate errors, the stiffness mapping module uses the principles of robot differential kinematics and the camera optical center velocity calculation formula to process the synchronized joint encoder data and calculate the camera self-motion velocity vector. Based on the assumption that all objects in the scene are absolutely rigid, the theoretical background optical flow vector is calculated pixel by pixel using the camera self-motion velocity vector and the rigid body motion field projection formula. Further, pixel regions with depth values ​​greater than the static anchor point depth threshold are selected as the static anchor point set, and the error compensation vector for the entire image is calculated using the residual field fitting formula. By subtracting the theoretical background optical flow vector and the error compensation vector from the observed optical flow vector acquired by the binocular depth camera, the net environmental response optical flow containing only the actual deformation of the plants is decoupled.

[0010] Preferably, in the process of generating a 3D stiffness map, the stiffness mapping module establishes an octree voxel map, combines the geometric curvature extracted from point cloud geometric information with the dynamic response variance map calculated based on the net environmental response optical flow, and iteratively calculates the posterior probability of flexibility of the voxels using the stiffness probability update formula; based on the posterior probability of flexibility, the voxels are marked in space as either penetrable flexible obstacles or rigid obstacles. This step utilizes Bayesian inference to fuse 2D visual features into 3D space, providing a physical property basis for path planning.

[0011] Preferably, in the process of constructing the directional cost field, the path planning module statistically models the local point cloud distribution of voxels in the flexible region of the 3D stiffness map, constructs the local covariance matrix using the local covariance matrix calculation formula, and extracts the main texture direction. Based on this main texture direction, an anisotropic damping cost tensor is constructed as the directional cost field using the anisotropic damping tensor construction formula. This cost field constrains the movement cost along the plant fiber direction to be lower than the movement cost perpendicular to the cutting direction.

[0012] Preferably, in the process of searching for and generating the optimal penetration path, the path planning module uses the elastic band algorithm to search for the initial path in three-dimensional space and uses the total energy cost function formula for iterative optimization; by minimizing the directional damping energy term contained in the formula, which is composed of the path tangent vector and the anisotropic damping cost tensor, a penetration path that conforms to the plant growth texture is generated, thereby reducing physical damage to biological tissues.

[0013] Preferably, during the output of real-time interactive features, the interactive control module calculates the visual optical flow divergence field, which characterizes the physical approach rate of the target surface, based on the net response optical flow of the environment during the execution process and using the optical flow field divergence calculation formula; based on the rigid body dynamics model of the six-degree-of-freedom robotic arm, it calculates the external contact torque, which characterizes the pure external collision load, using the contact torque observation formula; and outputs the visual optical flow divergence field and the external contact torque in combination.

[0014] Preferably, during the generation of the final control execution command, the industrial control computer compares the external contact torque and the visual optical flow divergence field with preset safe contact torque thresholds and visual divergence thresholds, respectively. When the external contact torque is less than the safe contact torque threshold and the visual optical flow divergence field is greater than the visual divergence threshold, it is determined to be an effective flexible penetration state; when the external contact torque is greater than or equal to the safe contact torque threshold and the visual optical flow divergence field is less than or equal to the visual divergence threshold, it is determined to be a rigid blocking state; when the visual optical flow divergence field is greater than the visual divergence threshold but the external contact torque approaches zero, it is determined to be a false contact state. The visual divergence threshold is 0.3 to 0.5; the safe contact torque threshold is 20% to 30% of the rated output torque of the joint motor.

[0015] Preferably, the industrial control computer generates joint drive command torque based on the judgment result using the adaptive impedance control torque calculation formula. When the condition is determined to be an effective flexible penetration state, the adaptive stiffness matrix parameter in the formula is reduced and the adaptive damping matrix parameter is increased, adjusting the robotic arm to a low stiffness characteristic to conform to the blade force; when the condition is determined to be a rigid blocking state, the emergency stop replanning logic is triggered and the high stiffness parameter is restored; when the condition is determined to be a false contact state, the adaptive stiffness matrix parameter and the adaptive damping matrix parameter remain unchanged, and the current motion trajectory continues to be executed.

[0016] This invention provides a method for 3D reconstruction and obstacle avoidance path planning for a weeding robotic arm based on binocular vision. It has the following beneficial effects: 1. This invention solves the problem that passive vision cannot distinguish environmental physical attributes by using an active excitation and motion decoupling mechanism. The system uses the mean of optical flow to trigger high-frequency micro-amplitude vibration, and combines kinematic compensation and residual fitting to eliminate camera self-motion interference, accurately extracting the net dynamic response of plants to excitation. This active perception method can quantitatively distinguish between flexible leaves and rigid stems, and construct a three-dimensional map containing physical stiffness information, providing environmental attribute basis for the obstacle avoidance decision of the weeding robot in complex vegetation environment.

[0017] 2. This invention constructs an anisotropic cost field based on the main direction of texture, which changes the traditional obstacle avoidance planning strategy of treating all obstacles as rigid no-go zones. By statistically analyzing the point cloud distribution of flexible regions, texture features are extracted, and a tensor field with low cost along the direction of plant fibers and high cost perpendicular to the cutting direction is established to guide the robotic arm to cut along the gaps in leaf growth. This compliant planning ensures that the robotic arm can penetrate the obstructed area to reach the target position while reducing the physical resistance and mechanical damage to the crop during the movement process.

[0018] 3. This invention establishes a real-time interactive and variable impedance control mechanism for visual-touch fusion, improving operational safety in unstructured environments. By jointly analyzing the visual optical flow divergence field and the external contact torque based on a momentum observer, the system can determine three states in real time: effective flexible penetration, rigid obstruction, and spurious contact. Based on this, the impedance parameters of the robotic arm are dynamically adjusted, reducing stiffness to adapt to the environment during flexible contact and triggering an emergency stop during rigid collisions, achieving a balance between operational efficiency and system safety in dense vegetation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the hardware architecture of a binocular vision-based 3D reconstruction and weeding robotic arm obstacle avoidance path planning method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the system logic module structure according to an embodiment of the present invention; Figure 3 This is a flowchart of a method for obstacle avoidance path planning of a 3D reconstruction and weeding robotic arm based on binocular vision, according to an embodiment of the present invention. Figure 4 This is the environmental dynamic characteristic monitoring and active stimulus response diagram of the present invention; Figure 5 This is a comparison diagram of the interaction force response characteristics at the end of the penetration process of the present invention; Figure 6 This is a statistical chart showing the success rate of operations under different complexity scenarios according to the present invention.

[0020] Among them, 10 is the mobile chassis; 20 is the six-degree-of-freedom robotic arm; 30 is the end effector; 40 is the binocular depth camera; 50 is the industrial computer; 100 is the stiffness mapping module; 200 is the path planning module; and 300 is the interactive control module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1The method provided in this embodiment of the invention operates on a robot system comprising a mobile chassis 10, a six-degree-of-freedom robotic arm 20, an end effector 30, a binocular depth camera 40, and an industrial control computer 50. The mobile chassis 10 is used to support system movement. The six-degree-of-freedom robotic arm 20 is mounted on the mobile chassis 10, and its joint motors are equipped with current feedback interfaces, supporting kilohertz-level torque data reading and position command response. The end effector 30 is fixed to the end of the six-degree-of-freedom robotic arm 20. The binocular depth camera 40 is rigidly connected to the end flange of the six-degree-of-freedom robotic arm 20 in an eye-to-hand manner, employing a global shutter sensor to eliminate image distortion under high-frequency vibration. The industrial control computer 50 is connected to the mobile chassis 10, the six-degree-of-freedom robotic arm 20, and the binocular depth camera 40 via an EtherCAT real-time industrial bus, performing data acquisition and command issuance.

[0023] See attached document Figure 2 The logic processing unit integrated within the industrial computer 50 includes: The stiffness mapping module 100 is used to process visual optical flow data and robotic arm kinematic data to construct a 3D map containing environmental physical properties; the path planning module 200 is used to generate a penetration path based on environmental texture features and stiffness distribution; and the interaction control module 300 is used to perform visual-touch data fusion and safety decisions during the execution phase.

[0024] See attached document Figure 3 This invention provides a method for obstacle avoidance path planning of a 3D reconstruction and weeding robotic arm based on binocular vision, comprising the following steps: S1, the industrial control computer 50 monitors the global optical flow data collected by the binocular depth camera 40. When it is determined that the average optical flow is lower than the preset static threshold, it injects a high-frequency, low-amplitude sinusoidal position disturbance command into the six-degree-of-freedom robotic arm 20 through the Ether CAT bus to generate an active excitation vibration signal. S2, the stiffness mapping module 100 receives the active excitation vibration signal and the synchronous joint encoder data, calculates the theoretical background optical flow field and combines the static anchor points in the image to eliminate the nonlinear error caused by the transmission gap of the robotic arm, extracts the environmental net response optical flow, and combines the point cloud geometric information to generate a three-dimensional stiffness map containing the distribution of rigid stems and flexible leaves. S3, the path planning module 200 constructs a direction cost field based on the main direction of the texture of the flexible region in the three-dimensional stiffness map, and searches and generates an optimal penetration path in the three-dimensional space based on the principle of energy optimization, which is parallel to the texture direction of the plant leaf. S4, the six-degree-of-freedom robotic arm 20 executes the optimal penetration path, and the interactive control module 300 calculates the visual optical flow divergence field of the contact area and the external contact torque based on the momentum observer in real time, and outputs real-time interactive features characterizing the current contact state. S5, the industrial control computer 50 determines the current state as effective flexible penetration, rigid blocking or false contact based on the real-time interaction characteristics, and adjusts the impedance control parameters of the six-degree-of-freedom robotic arm 20 or triggers emergency stop replanning to generate the final control execution command.

[0025] The implementation details of steps S1 to S5 will be described in detail below with reference to specific embodiments.

[0026] See attached document Figure 3 Regarding the specific implementation process of environmental dynamic detection and active stimulus triggering in step S1, this embodiment adopts the following logic for processing.

[0027] In unstructured agricultural scenarios, to accurately extract the physical stiffness characteristics of plants, the system must first eliminate background motion interference caused by natural wind or chassis vibration. The industrial computer 50 analyzes the visual flow field characteristics in real time, triggering active detection only when the environment is relatively stable, and generating excitation signals that conform to plant dynamics. The specific execution process is divided into the following sub-steps: S101, Image Sequence Acquisition and Preprocessing. The industrial computer 50 controls the binocular depth camera 40 to continuously acquire images at a fixed frame rate (e.g., 30fps), and performs Gaussian smoothing on the raw images to suppress sensor thermal noise. Subsequently, a dense optical flow algorithm (such as the Farneback algorithm) is used to calculate the instantaneous motion vector of each pixel on the image plane. To avoid misleading environmental state determination due to occlusion by the six-DOF robotic arm 20 or image edge distortion, this embodiment pre-constructs a region of interest mask, retaining only pixels within the effective area at the center of the field of view for subsequent calculations.

[0028] S102, Quantifying Environmental Dynamics. Since single-point optical flow is easily affected by texture reflections, the system uses statistical characteristics to characterize the overall stability of the environment. The industrial control computer 50 calculates the global optical flow mean, reflecting the intensity of motion across the current field of view, based on the collected optical flow field data using the global optical flow mean calculation formula. This step aims to establish a quantitative environmental state index; the global optical flow mean calculation formula is as follows: ; in, express The mean global optical flow at time t; This represents the set of pixels in an image defined by the region of interest mask; Represents a set Total number of effective pixels within (when When the value approaches zero, the system automatically skips the calculation of the current frame. Represents the pixel coordinates of the image plane; express The time is located at coordinate ( The two-dimensional optical flow vector at that location; This represents the Euclidean norm of a vector.

[0029] S103, Environmental Steady-State Determination. To prevent misjudgment caused by momentary camera shake, this embodiment introduces a time window verification mechanism. The industrial control computer 50 continuously monitors the global optical flow average, and only when the steady-state condition is stable... Frames (e.g.) The system is deemed to meet the conditions for active detection only when the global optical flow average of the binocular depth camera 40 is below a preset static threshold. The static threshold is preferably between 0.3 and 0.8 pixels per frame, calibrated based on the average noise level of the binocular depth camera 40 in a stationary state. If the determination passes, the system automatically enters active excitation mode; otherwise, the system maintains passive observation to avoid introducing ineffective excitation in strong wind environments.

[0030] S104, Generate a resonant matching excitation signal. To maximize the visual deformation response of the plant while ensuring the stability of the six-DOF robotic arm 20, this embodiment designs a frequency-matched sinusoidal excitation strategy. The industrial computer 50, based on pre-determined crop physical characteristics, uses a sinusoidal position perturbation formula to generate an active excitation vibration signal in Cartesian space. The sinusoidal position perturbation formula is: ; in, express Active excitation vibration signals are constantly applied to the end effector of the robotic arm; This represents the perturbation amplitude vector, with its magnitude set to 2mm to 5mm, and its direction preferably set to be parallel to the camera optical axis to enhance the dynamic changes of depth information. This indicates the preset crop resonance frequency, which is set to 2Hz to 5Hz for general broadleaf weeds; This represents a smooth envelope function used to linearly increase the amplitude during the initial excitation phase (e.g., the first 0.5 seconds) to prevent a step signal from causing harmful shocks to the 20 joints of the six-DOF robotic arm.

[0031] S105, Safety Injection and Execution. The industrial control computer 50 superimposes the calculated active excitation vibration signal into the current trajectory planning command and sends it to the six-DOF robotic arm 20 via the EtherCAT real-time industrial bus. During this process, the system monitors the condition number of the robotic arm's Jacobian matrix in real time. If a singular configuration risk is detected, the excitation signal injection is immediately blocked. In response to this high-frequency position command, the six-DOF robotic arm 20 drives the binocular depth camera 40 and its end effector to generate micro-vibrations, thereby artificially creating forced vibration characteristics independent of the natural environment in the visual image, providing the necessary physical excitation source for the subsequent analysis by the stiffness mapping module 100.

[0032] Regarding the specific implementation process of stiffness map construction based on kinematic decoupling and residual compensation in step S2, this embodiment adopts a strategy combining physical model-driven and data compensation, aiming to accurately extract the net response of plants to actively excited vibration signals from complex dynamic backgrounds. The stiffness mapping module 100 performs the calculation through the following rigorous logical flow: S201, Spatiotemporal alignment of multi-source data and camera motion calculation.

[0033] When the six-DOF robotic arm 20 performs high-frequency detection actions, the phase difference between the visual acquisition cycle and the joint control cycle makes direct calculation prone to introducing spurious optical flow. Therefore, the stiffness mapping module 100 employs a timestamp-based linear interpolation algorithm to strictly align the motion state fed back by the joint encoder to the image exposure time of the binocular depth camera 40. Based on the aligned joint data, the system utilizes the principles of robot differential kinematics to map the angular velocity in joint space to the instantaneous velocity of the camera's optical center in Cartesian space. The stiffness mapping module 100 calculates the camera's self-motion velocity vector using the camera optical center velocity calculation formula, which is: ; in, This represents the camera's self-motion velocity vector, which includes a linear velocity component. and angular velocity components ; This represents the time-aligned joint angle vector; This represents the time-aligned joint angular velocity vector; The geometric Jacobian matrix representing the end flange of the robotic arm relative to the base coordinate system; This represents the adjoint transformation matrix, which is uniquely determined by the hand-eye calibration parameters of the binocular depth camera 40, from the end flange coordinate system to the camera optical center coordinate system.

[0034] S202, constructing the theoretical background optical flow field under the rigid body assumption.

[0035] To isolate the flexible deformation of plants, the expected optical flow must first be calculated assuming all objects in the scene are absolutely rigid. The stiffness mapping module 100 combines the spatial geometry information provided by the depth map with the pinhole camera imaging model to derive the theoretical displacement of pixels due to camera motion. The system uses the rigid body motion field projection formula to calculate the theoretical background optical flow vector pixel by pixel. The rigid body motion field projection formula is: ; in, Represents pixel coordinates ( Theoretical background optical flow vector at the location; This represents the spatial depth value of the pixel (to prevent calculation divergence, when...). When the distance is less than the minimum imaging distance or is an invalid value NaN, the theoretical optical flow at that point is forced to zero. , This represents the normalized coordinates of the pixel on the normalized imaging plane; , , This represents the projection of the linear velocity component onto the three axes of the camera coordinate system; , , This represents the projection of the angular velocity components onto the three axes of the camera coordinate system.

[0036] S203, Adaptive Fitting and Compensation for Nonlinear Transmission Errors.

[0037] Considering the inherent transmission error and flexible deformation of the harmonic reducer in the six-DOF robotic arm 20, theoretical values ​​calculated solely based on kinematic formulas often have residual deviations. This embodiment utilizes the static background in the scene as a "virtual reference." The stiffness mapping module 100 selects pixel regions with depth values ​​greater than the static anchor point depth threshold (e.g., 2.0 meters, representing the ground or distant background) as a set of static anchor points. Since the physical velocity of the anchor point region is zero, the difference between its observed optical flow and theoretical optical flow represents the system error. Assuming this error exhibits a low-frequency, smooth distribution on the image plane, the system uses a residual field fitting formula to calculate the error compensation vector for the entire image. The residual field fitting formula is: ; in, Representing coordinates Error compensation vector at the location; Represents a second-order positional feature vector; This represents the polynomial coefficient matrix obtained by solving the problem using the least squares method based on the static anchor point set. If the number of points in the static anchor point set of the current frame is less than the minimum number of points required for fitting (e.g., 10), the system will maintain the polynomial coefficient matrix of the previous frame to ensure continuity.

[0038] S204, Environmental Net Response Extraction and Variance Feature Mapping.

[0039] After the above decoupling steps, the stiffness mapping module 100 subtracts the theoretical background optical flow vector and error compensation vector from the observed optical flow vector calculated by the dense optical flow algorithm to obtain the net environmental response optical flow vector. This vector only includes the actual deformation of the plant caused by forced vibration. To quantify flexibility, the system... For example, the variance of the net response modulus is calculated within 10 frames to generate a dynamic response variance map. The larger the variance value, the greater the deformation amplitude of the object under the same excitation, that is, the lower the physical stiffness.

[0040] S205, Update of 3D stiffness map based on Bayesian inference.

[0041] To support subsequent path planning, 2D visual features need to be fused into 3D space. The stiffness mapping module 100 establishes an octree voxel map, combining the geometric curvature of the point cloud (morphological features distinguishing leaves from stems) and the dynamic response variance map (physical features distinguishing flexible from rigid), and updates voxel attributes using a log-odds Bayesian filter. The stiffness mapping module 100 iteratively calculates the posterior probability of flexibility for voxels using a stiffness probability update formula, which is: ; in, Indicates time No. The flexible posterior probability that an individual element belongs to a flexible region; This represents the prior flexible probability at the previous moment; This represents the flexible likelihood probability based on the current observation data. This probability is obtained by normalizing the dynamic response variance of the pixels projected onto the voxel using the Sigmoid function. Finally, the industrial control computer 50 marks the voxel as a "penetrable flexible obstacle" or a "rigid obstacle" in the 3D stiffness map based on whether the probability value corresponding to the flexible posterior probability exceeds 0.8.

[0042] Regarding the specific implementation process of compliant path planning based on anisotropic flow fields in step S3, this embodiment no longer treats the environment as a binary opposition of "obstacles" and "free space." Instead, it utilizes the path planning module 200 to construct a mathematical field containing directional constraints, guiding the six-degree-of-freedom robotic arm 20 to compliantly penetrate through gaps in the vegetation. The specific execution process is as follows: S301, texture principal direction extraction based on local point cloud statistics.

[0043] To ensure the robot's penetration motion conforms to the natural growth texture of plant leaves or stems, thereby minimizing physical damage to biological tissues, the path planning module 200 performs geometric feature analysis on voxels marked as "penetrable flexible regions" based on the 3D stiffness map generated in step S2. The system assumes that the direction with the least change in local normal vector on the plant surface is the texture growth direction, and uses principal component analysis to statistically model the local point cloud distribution. The path planning module 200 traverses the voxels of the flexible region, searching for each voxel... Neighborhood (e.g., taking) The point cloud is within the range of 20 to 50, and the local covariance matrix is ​​constructed using the local covariance matrix calculation formula. The local covariance matrix calculation formula is: ; in, This represents the local covariance matrix at the current voxel point; Indicates the number of neighboring points; Represents the first in the neighborhood A point cloud coordinate vector; This represents the geometric center coordinate vector of all points within this neighborhood. (Path planning module 200 pairs) Perform eigenvalue decomposition, select the eigenvector corresponding to the smallest eigenvalue as the surface normal, and select the eigenvector corresponding to the largest eigenvalue as the principal texture direction. (Needs to be normalized to a unit vector). If the number of valid points in the neighborhood is less than 3, the system marks that location as an isotropic region, i.e. Let it be the zero vector.

[0044] S302, construct an anisotropic damping cost field.

[0045] To mathematically express the physical property of "low resistance along the grain, high resistance against the grain" in the path search space, this embodiment constructs a second-order tensor field instead of a traditional scalar cost map. Based on the extracted principal direction of the texture, the path planning module 200 constructs an ellipsoidal damping model for each discrete point in the space, compressing the cost of movement along the direction of the plant fiber while amplifying the cost perpendicular to the direction of fiber cutting. The path planning module 200 calculates the anisotropic damping cost tensor using the anisotropic damping tensor construction formula, which is: ; in, Indicates the location in spatial coordinates Anisotropic damping cost tensor at the location; Represents the identity matrix; This indicates the main direction of the texture at that coordinate. This represents the tangential damping coefficient, with a value range of 0.1 to 0.5, designed to reduce the algorithmic penalty for conforming to texture movement; The normal damping coefficient is set to a value between 5.0 and 10.0, designed to suppress the generation of paths that penetrate vertically through the blade. This tensor field forms the basis for the directional constraints of subsequent path optimization.

[0046] S303, the optimal penetration path generation based on the principle of minimum energy.

[0047] After obtaining the directional constraints, the path planning module 200 uses the Elastic Band algorithm to search for the optimal solution connecting the initial and target configurations in three-dimensional space. This process treats the path as an elastic rope subjected to multiple force fields, achieving equilibrium by minimizing the system's total energy functional. The system first generates a collision-free initial coarse path using the Fast Random Search Tree (RRT*) algorithm, then discretizes it and iteratively optimizes it using the total energy cost function formula. The total energy cost function formula is: ; in, This represents the total energy cost of the planned path; This represents the total number of discrete path points; Indicates the first A vector of coordinates of path points; This represents the smoothing weight coefficient, used to constrain path length and curvature, ensuring the continuity of motion of the six-DOF robotic arm 20. This represents the stiffness obstacle avoidance weight coefficient; This represents the repulsive potential field value of a rigid obstacle obtained from a table based on a 3D stiffness map. For rigid stem regions, this value tends towards infinity; the last term of the formula is the directional damping energy term, derived from the path tangent vector ( With anisotropic damping cost tensor The quadratic form is constructed. The path planning module 200 uses the Levenberg-Marquardt nonlinear optimization algorithm to solve the above objective function until the energy converges, and finally outputs an optimal penetration path that avoids the rigid support structure and cuts parallel to the texture of the plant leaves.

[0048] Regarding the specific implementation process of real-time interactive feature calculation for visual-touch fusion in step S4, this embodiment addresses the perception blind spots caused by the single visual modality's susceptibility to occlusion under foliage and the single force modality's lack of response to non-contact states during the execution of the penetration path by the six-DOF robotic arm 20. The interaction control module 300 synchronously and in parallel processes the geometric deformation features of the visual field and the dynamic response features of the joint space, outputting high-dimensional real-time interactive features. The specific execution process is as follows: S401, Extraction of differential features of visual divergence in the contact area.

[0049] As the robotic arm's end effector approaches and attempts to push aside obstructing branches and leaves, the optical flow field on the image plane exhibits a significant central divergence. To quantify this visual expansion effect caused by the shortened approach distance and surface deformation, the interactive control module 300, based on the divergence principle in fluid mechanics, identifies the central region of the image (i.e., directly in front of the end effector) as the region of interest for end-effector contact. The system reuses the net environmental response optical flow vector (with camera self-motion components removed) calculated in step S2 and uses a discretized Green's Theorem to calculate the flux density of this flow field within the region of interest for end-effector contact, thus characterizing the physical approach rate to the target surface. The interactive control module 300 uses the optical flow field divergence calculation formula to calculate the statistical characteristic values ​​of the visual optical flow divergence field in the contact area. The optical flow field divergence calculation formula is: ; in, The value represents the visual interaction feature, and its physical meaning is the intensity of the optical flow source per unit area. Positive values ​​represent expansion (approaching / pushing away), and negative values ​​represent contraction (moving away). This represents the set of pixels that touch the region of interest at the end. This represents the total number of valid pixels in the area. If this value is lower than a preset statistical threshold (e.g., 50 pixels), the system will forcibly remove the pixel count. Set to zero to prevent division by zero errors and noise amplification. and Representing pixel coordinates The horizontal and vertical components of the net optical flow vector of the environment; and This represents the spatial partial derivative of the optical flow vector on the image plane. In this embodiment, the Sobel operator is used for convolution calculation to obtain a more robust gradient estimate.

[0050] S402, external contact torque decoupling based on momentum observer.

[0051] While visual features can provide early warning of pre-contact trends, they are highly susceptible to interference from changes in light intensity and leaf shading in high-canopy environments. To obtain accurate physical contact force feedback while avoiding the introduction of expensive and easily damaged end-effector six-dimensional force sensors, this embodiment constructs a generalized momentum observer based on the rigid body dynamics model of the six-DOF manipulator 20. This method utilizes the time integration characteristics of the manipulator's momentum to effectively avoid the high-frequency noise amplification problem caused by directly differentiating the acceleration from the joint position. The interactive control module 300 collects joint motor current and position data in real time and iteratively calculates the external contact torque based on the momentum observer using the contact torque observation formula. To achieve digital discrete calculation, the system uses the observation values ​​from the previous control cycle for integral updates. The contact torque observation formula is: ; in, Indicates time The estimated vector of external contact torque reflects the pure external collision loads on each joint of the robotic arm; This represents the observer gain matrix, which is usually selected as a positive definite diagonal matrix with diagonal elements between 15.0 and 30.0. This parameter determines the bandwidth of the observer. If the value is too large, it will introduce noise, and if the value is too small, it will cause response lag. The inertia matrix represents the robotic arm; and These represent the joint angle vector and the joint angular velocity vector, respectively. This represents the joint driving torque vector calculated through motor current feedback. This represents the transpose of the Coriolis force and centrifugal force matrices. The transpose term is introduced here to utilize the equations of robot dynamics. The skew-symmetry eliminates the acceleration term; Represents the gravity compensation vector; This represents the generalized momentum of the system at the initial moment; This represents the integral variable.

[0052] After obtaining the above two sets of features, the interactive control module 300 performs nearest neighbor alignment based on timestamps between the acquisition time of the visual data and the calculation time of the torque data. Characterizing deformation trend) and The combined output (characterizing contact force) is a real-time interactive feature. Based on this feature, the industrial computer 50 can accurately determine in the subsequent S5 step whether the current state is "flexible blade brushing past" (high light flow, low contact torque) or "rigid branch collision" (low light flow, high contact torque).

[0053] Regarding the specific implementation process of the safety closed-loop decision-making and control based on visual-touch association in step S5, after acquiring the high-dimensional interactive features output in S4, the interactive control module 300 needs to establish a logical connection between visual morphological changes and mechanical feedback in order to achieve both compliant and safe penetration operations in an unstructured and dynamically changing plant environment. This embodiment, based on the physical principle that "flexible biological tissues are easily deformed under pressure, while rigid support structures are difficult to displace under pressure," constructs a multi-state arbitration mechanism and dynamically adjusts the motion compliance of the six-degree-of-freedom robotic arm 20 in the control loop. The specific execution process is as follows: S501 performs multi-state correlation arbitration based on visual divergence and contact torque.

[0054] To accurately map abstract sensor data to physical interaction scenarios, the interaction control module 300 presets two key physical boundary parameters: visual divergence threshold. For example, the value can be between 0.3 and 0.5. This value is calculated based on the geometric projection relationship between the camera's field of view and the typical blade size, representing significant texture spread and the safe contact torque threshold. For example, it can be set to 20% to 30% of the rated output torque of the joint motor to characterize the safe load boundary. The visual interaction feature values ​​output by the interactive control module 300 according to step S4. Magnitude of the estimated vector of contact torque with external force The current penetration state is divided into three typical operating conditions using multi-dimensional logic thresholds: First, when the system detects and At this point, it is determined to be in an "effective flexible penetration state". At this time, obvious light flow divergence is detected visually (the branches and leaves are pushed aside), and the contact force remains at a low level, indicating that the six-degree-of-freedom robotic arm 20 is correctly cutting into the gaps in the leaf texture, and the physical interaction conforms to the flexible characteristics.

[0055] Secondly, when the system detects and When the object is in a "rigid obstruction state", the contact force rises rapidly and approaches the safety boundary, but the visual optical flow field does not expand significantly (the field of vision is filled with stationary obstacles or the optical flow field is chaotic and disordered), indicating that the end has encountered a thick branch or facility support that cannot be pushed away, and continuing to advance will lead to mechanical damage.

[0056] Third, when the system detects but When the noise level is significantly lower than the contact noise baseline, it is considered a "false contact state." This situation is usually caused by wind blowing the blades or drastic changes in light and shadow, resulting in a visual false alarm. In this case, the robotic arm has not made any substantial physical contact, and the system should ignore the visual disturbance and maintain the original motion trajectory.

[0057] S502, adaptive adjustment of impedance parameters and emergency stop strategy in response to interactive states.

[0058] In response to the state determination result in S501 above, the interactive control module 300 no longer executes high-stiffness position servo control, but switches to a variable-parameter impedance control strategy in Cartesian space. This strategy adapts to the environment by constructing a "virtual spring-damped" system between the end effector of the robotic arm and the environment, and dynamically reconstructs the mechanical impedance characteristics of the system. The industrial computer 50 adjusts the impedance parameters in real time based on the current arbitration state, and uses the adaptive impedance control torque calculation formula to generate the final joint drive command torque issued to the six-degree-of-freedom robotic arm 20. The adaptive impedance control torque calculation formula is as follows: ; in, This indicates the joint drive command torque sent to the motor driver; This represents a gravity compensation term based on the robot's dynamics model, used to offset the robot arm's own weight and ensure that it is in a state of gravitational balance when there are no commands. This represents the transpose of the Jacobian matrix, whose physical function is to losslessly map virtual elastic forces in Cartesian space to torques in joint space (note that the path planning layer has already avoided this). The singular configuration ensures the reversibility of the force mapping. and Let them represent the expected Cartesian position vector and the current Cartesian position vector at the end, respectively; and Let represent the expected Cartesian velocity vector and the current Cartesian velocity vector at the end, respectively.

[0059] In the formula and They are respectively determined by the state The adaptive stiffness matrix and adaptive damping matrix are variable (both are positive definite diagonal matrices). The specific adaptive adjustment strategy is as follows: If determined to be in an "effective flexible penetration state," the system will significantly reduce... The values ​​of the main diagonal elements (e.g., reduced to 20% of the rated stiffness) and simultaneously increased. (For example, by increasing the damping to 150% of the rated damping), the robotic arm exhibits the physical characteristics of "low stiffness and high damping", thus enabling it to deviate slightly in response to the pushing force of the blades without generating a destructive force. If the condition is determined to be a "rigid blocking state", the industrial control computer 50 will immediately trigger the emergency stop and replanning logic, freezing the system. The update (i.e., stop trajectory interpolation) and restore high stiffness to maintain the current attitude, avoid the robot arm from becoming unstable due to excessive pushing, and at the same time report the blockage information to the path planning module 200 to request a new path search; If the system determines the contact status as "false," it will maintain the current status. and With the parameters unchanged, the original trajectory will continue to be executed to prevent motion jitter caused by visual noise.

[0060] Specific application examples: This embodiment applies the binocular vision-based 3D reconstruction and obstacle avoidance path planning method for weeding robotic arms proposed in this invention to a high-canopy-density cornfield scenario, focusing on verifying the system's ability to identify and handle both flexible leaf occlusion and rigid stem obstruction.

[0061] Experiment initialization and environment configuration: The experimental setup employed a wheeled mobile platform equipped with a six-DOF robotic arm 20. The end effector 30 was a touch-sensitive mechanical weeding blade. A binocular depth camera 40 was mounted on the end flange of the robotic arm. The experimental environment was a mid-growth cornfield with plant spacing of 25-35 cm and abundant weeds. A static detection threshold of 0.6 pixels / frame and a sinusoidal position perturbation frequency were set. The maximum disturbance amplitude is 4Hz. The value is 4mm. In the path planning module 200, the ratio of tangential damping to normal damping coefficient is set to 1:40.

[0062] Experimental procedure and numerical analysis: Environmental steady-state monitoring and excitation injection. An industrial computer (ICC50) acquires image sequences and calculates the initial average global optical flow of the environment. Pixels / frame. Due to The system determines the environment to be in a steady state and triggers active detection. Based on the sinusoidal position perturbation formula, in At the (positive peak point), the applied end displacement offset is: ; The excitation caused the binocular depth camera (40) to generate forced vibrations, and the dynamic response characteristics of the plant were obtained.

[0063] Stiffness map construction and path search. The stiffness mapping module 100 extracts the net environmental response optical flow, obtaining a variance eigenvalue of 0.85 at the corn leaf and 0.08 at the stem. A Bayesian filter then marks the leaves as permeable regions. The path planning module 200 uses PCA to extract the principal direction of the leaf texture. Guided by an anisotropic flow field, the generated path, when approaching the blade, has a tangential movement cost that is much lower than a normal cutting cost, thus achieving parallel-texture penetration.

[0064] Real-time interaction and adaptive control. The interactive control module 300 performs real-time calculations as the robotic arm executes the penetration path. At a certain moment, the external contact torque vector magnitude is observed to be 1.8 N·m, and the visual divergence eigenvalue... The value is 0.72. Numerical judgment logic: At this point, the condition is met. and The industrial control computer confirmed that the 50-degree flexible penetration was effective. Impedance adjustment calculation: set the initial stiffness. After adjustment If the end of the device deviates by 2 cm due to blade resistance, the feedback torque is reduced from 8 N to 1.6 N, ensuring low-damage contact with the crop.

[0065] in conclusion: See attached document Figures 4-6 Based on the data in the figure and the technical features of this solution, the following conclusions can be drawn: Active stimulation effectively solves the problem of missing physical characteristics in complex environments. (Refer to...) Figure 4 In unstructured agricultural scenarios, when the mean global optical flow of the environment is below the static judgment threshold of 0.6, the system successfully induced the dynamic response of the target by injecting sinusoidal position perturbations. This mechanism enables the robotic arm to actively detect the physical properties of plants and acquire pixel-level deformation data even in a static environment, providing the necessary prerequisite for distinguishing between rigid stems and flexible leaves.

[0066] The view-touch fusion strategy significantly reduces collision load during interaction. According to Figure 5 A comparison of the torque curves reveals that, during penetration tasks, the comparative scheme (artificial potential field method), lacking classification and arbitration of the interaction state, experienced a rapid increase in its contact torque modulus, exceeding the safe contact torque threshold of 5.0, posing a risk of damaging crops or mechanical structures. In contrast, the present invention combines visual optical divergence and momentum observers, enabling real-time determination of the interaction state and dynamically adjusting the adaptive stiffness matrix to maintain the torque at approximately 1.8 N·m. This result demonstrates that the impedance-adaptive strategy enables the robotic arm to exhibit low stiffness and high damping characteristics, achieving compliant contact with flexible biological tissues.

[0067] Anisotropic path planning significantly improves operational reliability in dense environments. (Refer to...) Figure 6 Statistical data shows that in highly dense shading environments, this solution achieved a success rate of 92.8%, significantly higher than the 72.4% of the comparative solution. This is mainly attributed to the directional cost field constructed in step S3, which guides the robotic arm along the safe range defined by datal, penetrating the obstacles by following the leaf texture. Experimental data confirms that this solution, while maintaining operational efficiency, significantly enhances the survivability and operational accuracy of the weeding robot in unstructured obstacle groups by utilizing gaps in plant growth to avoid rigid obstacles.

[0068] In summary, the present invention solves the problems of traditional obstacle avoidance algorithms having a single perception dimension and excessively rigid operation in agricultural scenarios by deeply coupling active perception and compliant control.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for obstacle avoidance path planning of a 3D reconstruction and weeding robotic arm based on binocular vision, characterized in that, Includes the following steps: The industrial control computer (50) monitors the global optical flow data collected by the binocular depth camera (40). When it is determined that the average global optical flow is lower than the preset static judgment threshold, it injects a sinusoidal position disturbance command into the six-degree-of-freedom robotic arm (20) mounted on the mobile chassis (10), which drives the end effector (30) to generate an active excitation vibration signal. The stiffness mapping module (100) receives the active excitation vibration signal and the synchronized joint encoder data, eliminates transmission error and extracts the net response optical flow of the environment, and generates a three-dimensional stiffness map by combining point cloud geometric information. The path planning module (200) constructs a directional cost field based on the main texture direction of the flexible region in the three-dimensional stiffness map, and searches for and generates the optimal penetration path in three-dimensional space; The six-degree-of-freedom robotic arm (20) executes the optimal penetration path, and the interactive control module (300) calculates the visual light flow divergence field and external contact torque in real time, and outputs real-time interactive features. The industrial control computer (50) determines the current state as an effective flexible penetration state, a rigid blocking state, or a false contact state based on the real-time interaction features, and generates the final control execution command accordingly.

2. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of the industrial control computer (50) monitoring the global optical flow data and generating the active excitation vibration signal: The industrial control computer (50) processes the collected global optical flow data using the global optical flow mean calculation formula to obtain the global optical flow mean that reflects the intensity of the overall motion of the environment. When the global optical flow mean is lower than the preset static judgment threshold, the industrial control computer (50) uses a sinusoidal position perturbation formula containing a smooth envelope function and a preset crop resonance frequency to generate the sinusoidal position perturbation command in Cartesian space. The six-degree-of-freedom robotic arm (20) responds to the sinusoidal position disturbance command and drives the end effector (30) to generate the active excitation vibration signal; The static determination threshold is 0.3 to 0.8 pixels per frame.

3. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of the stiffness mapping module (100) receiving the joint encoder data and eliminating transmission errors: The stiffness mapping module (100) uses the principle of robot differential kinematics and the camera optical center velocity calculation formula to process the synchronized joint encoder data and calculate the camera self-motion velocity vector. Based on the assumption that all objects in the scene are absolutely rigid, the theoretical background optical flow vector is calculated pixel by pixel using the camera's self-motion velocity vector and the rigid body motion field projection formula.

4. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 3, characterized in that, In the step of the stiffness mapping module (100) extracting the net environmental response optical flow and generating the three-dimensional stiffness map: The stiffness mapping module (100) selects pixel regions with depth values ​​greater than the static anchor point depth threshold as a set of static anchor points, and uses the residual field fitting formula to calculate the error compensation vector of the whole image. Subtract the theoretical background optical flow vector and the error compensation vector from the observed optical flow vector acquired by the binocular depth camera (40) to obtain the net environmental response optical flow that only contains the actual deformation of the plant.

5. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of generating the three-dimensional stiffness map by combining the point cloud geometric information in the stiffness mapping module (100): The stiffness mapping module (100) establishes an octree voxel map, combines the geometric curvature extracted from the point cloud geometric information with the dynamic response variance map calculated based on the net environmental response optical flow, and uses the stiffness probability update formula to iteratively calculate the flexible posterior probability of the voxel. Based on the flexible posterior probability, voxels are marked in space as either permeable flexible or rigid obstacles to form the three-dimensional stiffness map.

6. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of the path planning module (200) constructing the directional cost field based on the three-dimensional stiffness map: The path planning module (200) performs statistical modeling of the local point cloud distribution of the voxels in the flexible region of the three-dimensional stiffness map, constructs the local covariance matrix using the local covariance matrix calculation formula, and extracts the main direction of the texture. Based on the main direction of the texture, an anisotropic damping cost tensor is constructed using the anisotropic damping tensor construction formula, and the anisotropic damping cost tensor is used as the directional cost field, so that the movement cost along the plant fiber direction is lower than the movement cost perpendicular to the cutting direction.

7. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 6, characterized in that, In the step of the path planning module (200) searching and generating the optimal penetration path: The path planning module (200) uses the elastic band algorithm to search for an initial path in three-dimensional space and uses the total energy cost function formula for iterative optimization; The optimal penetration path is generated by minimizing the directional damping energy term, which is composed of the path tangent vector and the anisotropic damping cost tensor, included in the total energy cost function formula.

8. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of the interactive control module (300) calculating the visual optical flow divergence field and the external contact torque and outputting the real-time interactive features: The interactive control module (300) calculates the visual optical flow divergence field, which characterizes the physical approach rate of the target surface, based on the net response optical flow of the environment during the execution process and using the optical flow field divergence calculation formula. The interactive control module (300) calculates the external contact torque characterizing the pure external collision load based on the rigid body dynamics model of the six-degree-of-freedom manipulator (20) using the contact torque observation formula; The real-time interactive features are output by combining the visual optical flow divergence field with the external contact torque.

9. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step of the industrial control computer (50) determining the current state based on the real-time interaction features: The industrial control computer (50) compares the external contact torque and the visual optical flow divergence field that constitute the real-time interactive features with the preset safe contact torque threshold and visual divergence threshold, respectively. When the external contact torque is less than the safe contact torque threshold and the visual optical flow divergence field is greater than the visual divergence threshold, it is determined to be the effective flexible penetration state; When the external contact torque is greater than or equal to the safe contact torque threshold and the visual optical flow divergence field is less than or equal to the visual divergence threshold, it is determined to be the rigid blocking state; When the visual optical flow divergence field is greater than the visual divergence threshold but the external contact torque approaches zero, it is determined to be a false contact state. The visual divergence threshold is 0.3 to 0.5; the safe contact torque threshold is 20% to 30% of the rated output torque of the joint motor.

10. The method for obstacle avoidance path planning of a three-dimensional reconstruction and weeding robotic arm based on binocular vision according to claim 1, characterized in that, In the step where the industrial control computer (50) generates the control execution command based on the determination result: The industrial control computer (50) generates joint drive command torque based on the judgment result using the adaptive impedance control torque calculation formula; When the effective flexible penetration state is determined, the adaptive stiffness matrix parameter in the adaptive impedance control torque calculation formula is reduced and the adaptive damping matrix parameter is increased to adjust the joint drive command torque. When the rigid blocking state is determined, the emergency stop replanning logic is triggered and the high stiffness parameters are restored to generate the final control execution command. When the false contact state is determined, the current adaptive stiffness matrix parameters and adaptive damping matrix parameters are kept unchanged, and the joint drive command torque is continued to be executed.

Images