Low-magnetic interference robot control system and method based on image recognition

Abstract

This invention relates to the field of robotic arm control technology, specifically providing a low-magnetic-interference robotic arm control system and method based on image recognition. It aims to solve problems in existing robotic arm control systems, such as placement deviations affecting measurement accuracy and insufficient recognition capabilities of general image recognition algorithms leading to placement misalignment. To this end, the low-magnetic-interference robotic arm control system based on image recognition of this invention includes: an image acquisition module for acquiring images of samples; an image recognition module for determining the spatial pose parameters and category identification information of the samples; and a control module for generating control commands based on the spatial pose parameters and category identification information, and controlling the robotic arm to execute the control commands. This invention, by determining the robotic arm's operating path based on spatial pose parameters, can accurately grasp and place samples, effectively improving placement accuracy.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, specifically providing a low-magnetic-interference robotic arm control system and method based on image recognition. Background Technology

[0002] With the deepening of research in fields such as geophysics, geological surveys, and magnetic analysis of materials, magnetic measurement equipment is widely used in the automated testing of rock samples, mineral materials, and high-sensitivity magnetic parameters. Typical magnetic measurement equipment, such as magnetometers and magnetic susceptibility meters, are characterized by high sensitivity, wide measurement range, and high sensitivity to external interference.

[0003] In traditional experiments, the placement and replacement of test samples are highly dependent on manual operation. The specific process includes: manually identifying the sample number, manually grabbing the sample, opening the detection chamber, placing the sample, turning off the equipment and triggering the measurement. The operation is frequent and requires high repeatability, resulting in low detection efficiency and easy positional deviation, which affects the measurement accuracy.

[0004] Currently, some industrial automation systems use six-axis robotic arms to complete loading and unloading tasks, improving inspection efficiency while reducing errors caused by human factors. However, traditional robotic arms typically use static path planning and lack real-time attitude correction and error compensation mechanisms, making them prone to placement deviations. In magnetic measurement scenarios, samples usually need to be placed in the detection chamber of the magnetic measurement equipment with high repeatability (e.g., ±0.05mm). If placement deviations occur, they will seriously affect the measurement accuracy.

[0005] When determining the operating path of a robotic arm, sample images are typically acquired, and their position is determined by image recognition, thus defining the robotic arm's path. Currently, common image recognition algorithms (such as YOLOv3 / v5, ResNet, etc.) are commonly used. However, due to the small size, smooth surface, and reflective material of samples in magnetic measurement scenarios, these general models are prone to recognition drift, false detections, and inability to extract posture information under complex laboratory backgrounds or non-uniform lighting conditions, leading to robotic arm grasping failures or placement deviations. Furthermore, existing algorithms often focus only on position detection, lacking the ability to extract geometric parameters such as posture (e.g., rotation angle, tilt direction), which is a serious weakness in high-precision placement scenarios and cannot meet the requirements of high-precision placement.

[0006] Accordingly, a new technical solution is needed in this field to solve the above problems. Summary of the Invention

[0007] The present invention aims to solve the above-mentioned technical problems, namely, to solve the problems in the existing technology of traditional robotic arm control systems that are prone to placement deviations that affect measurement accuracy and the problems of insufficient recognition capabilities of general image recognition algorithms that lead to placement offsets.

[0008] In a first aspect, the present invention provides a low-magnetic-interference robotic arm control system based on image recognition, comprising: an image acquisition module for acquiring images of magnetic samples; an image recognition module for recognizing the images acquired by the image acquisition module, determining the spatial pose parameters of the magnetic sample and the category identification information of the magnetic sample, wherein the spatial pose parameters include coordinates and rotation angles; and a control module for receiving the spatial pose parameters and the category identification information, generating control commands based on the spatial pose parameters and the category identification information, and controlling the robotic arm to execute the control commands.

[0009] In the preferred embodiment of the above control system, the image recognition module includes a reflection robust preprocessing submodule, a feature extraction submodule, a channel attention submodule, and a multi-task output submodule connected in sequence. The reflection robust preprocessing submodule performs edge enhancement and reflection suppression on the input image. The feature extraction submodule extracts multi-scale feature maps from the preprocessed image. The feature extraction submodule is constructed using a MagNet neural network, which includes a convolutional backbone and a multi-scale feature fusion structure. The convolutional backbone consists of multiple convolutional modules, at least one of which employs depthwise separable convolution. The channel attention submodule adaptively enhances the spatial dimensions of each channel of the multi-scale feature map. The multi-task output submodule synchronously outputs the spatial pose parameters and the category identification information based on the enhanced multi-scale feature map.

[0010] In the preferred embodiment of the above control system, the reflection robust preprocessing submodule is used to identify the bright areas in the input image and suppress the gradient magnitude of the pixels corresponding to the bright areas.

[0011] In the preferred embodiment of the above control system, the channel attention submodule includes: a global average pooling subunit, used to perform global average pooling on each channel of the multi-scale feature map in the spatial dimension, compressing the features of each channel into a scalar to generate a channel description vector; a weight generation subunit, including an activation function layer, a fully connected layer and a sigmoid function layer connected in sequence, used to generate a corresponding channel weight vector based on the channel description vector; and a channel weighting subunit, used to perform channel-level multiplication on the input multi-scale feature map and the channel weight vector to weight each channel.

[0012] By employing the aforementioned technical solutions, the image recognition module of this application combines a lightweight convolutional structure with a reflection robustness mechanism. Its recognition performance under complex backgrounds such as reflective interference, weak contrast, and stacked occlusion is significantly superior to conventional models like YOLO and ResNet, significantly improving image recognition accuracy and capture success rate, and demonstrating good engineering practicality and robustness. Furthermore, the feature extraction submodule uses a lightweight network structure with depthwise separable convolutions, maintaining a high recognition rate on low-computation platforms, outperforming conventional models like YOLO and ResNet.

[0013] In the preferred embodiment of the above control system, the image acquisition module is an RGB camera, a depth camera, or a binocular stereo camera.

[0014] In the preferred embodiment of the above control system, the arm body, transmission joints and / or grippers of the robotic arm are made of non-magnetic materials; and / or the end effector of the robotic arm is a hollow cup brushless DC motor.

[0015] In the preferred embodiment of the above control system, the robotic arm is equipped with a shield, which is at least located on the outside of the end effector of the robotic arm and / or the transmission joint of the robotic arm body.

[0016] By employing the above-mentioned technical solution, and utilizing non-magnetic materials to fabricate the arm body, transmission joints, and grippers of the robotic arm, and selecting a hollow cup brushless DC motor as the end effector to reduce residual magnetism and transient electromagnetic disturbances during operation, the problem of magnetic field interference to magnetic measurement equipment caused by traditional robotic arms is effectively solved. Furthermore, this application also provides shielding covers at least on the outside of the end effector and the transmission joints of the robotic arm body. These shielding covers are typically made of high-permeability materials, which can absorb leakage magnetism in a timely manner, preventing magnetic interference. Through optimized design of the robotic arm's fabrication materials, motor type, and shielding cover, this application can significantly reduce magnetic interference by 91.3%. In actual testing, measurements of the magnetic field disturbances generated during the robotic arm's operation revealed that, compared to traditional robotic arms, the amplitude of magnetic field disturbances generated by the robotic arm in this application at a typical working distance is reduced by approximately 91.3%.

[0017] In a preferred embodiment of the above control system, the robotic arm includes a robotic arm body and a gripper disposed at the end of the robotic arm body. A magnetic field sensor is disposed on the robotic arm body near the gripper. The magnetic field sensor is used to detect the magnetic field strength of the environment in which the gripper is located. The control system further includes an adjustment module, which is communicatively connected to the control module. The adjustment module includes a first adjustment submodule and a second adjustment submodule. The first adjustment submodule is used to adjust the extension length of the gripper based on the magnetic field strength. The second adjustment submodule is used to adjust the opening angle of the gripper based on the size of the magnetic sample and the magnetic field strength.

[0018] In the preferred embodiment of the above control system, the first adjustment submodule is used to adjust the extension length of the gripper to the minimum retraction length when the magnetic field strength is less than or equal to a first threshold, and to adjust the extension length of the gripper according to a preset function when the magnetic field strength is greater than the first threshold and less than or equal to a second threshold. The preset function is a linear function or a piecewise function, the first threshold is the upper limit of the safe magnetic field, and the second threshold is the upper limit of the magnetic field of the detection cavity of the magnetic measuring device.

[0019] When the above technical solution is adopted, the magnetic field strength of the environment in which the gripper is located is detected in real time by setting a magnetic field sensor to determine whether there are abnormal fluctuations or interference in the current magnetic field strength. Then, based on the magnetic field strength and / or the size of the magnetic sample, the extension length and opening angle of the gripper are adjusted to realize the dynamic extension and retraction of the gripper and the flexible adjustment of the gripping angle. This can avoid interference sources, optimize the gripping posture, and be compatible with gripping tasks of samples of various sizes and material types.

[0020] The technical effects that can be obtained by this invention are as follows:

[0021] (1) By using non-magnetic materials to prepare key parts such as the transmission joints, arm body and gripper of the robotic arm, and using a hollow cup brushless DC motor as the end effector, and setting a magnetic shielding structure on the outside of the transmission joints and the motor, the magnetic field interference generated during the operation of the robotic arm can be effectively reduced and the stability of the magnetic measurement environment can be improved.

[0022] (2) The image recognition module of this application adopts a combination of lightweight convolution structure and reflection robust preprocessing mechanism, which can achieve stable recognition of magnetic samples in the presence of reflection, complex lighting and stray background, and can simultaneously output the spatial pose parameters and category identification information of magnetic samples, thereby improving the accuracy of magnetic sample positioning and attitude recognition.

[0023] (3) By detecting the magnetic field strength of the environment where the gripper is located in real time, and adjusting the extension length and opening angle of the gripper based on the magnetic field strength and the size information of the magnetic sample, the robotic arm can adapt to the gripping needs of magnetic samples of different sizes and shapes, and improve the system's adaptability to various magnetic measurement scenarios.

[0024] (4) By combining the running path correction strategy with the feedback of the magnetic field sensor, the running path of the robotic arm can be dynamically adjusted, which can effectively improve the trajectory control accuracy during the magnetic sample grasping and placement process, enabling the robotic arm to achieve stable and safe placement of magnetic samples in a narrow detection cavity environment.

[0025] (5) The system adopts a modular structure design. The image recognition module, control module and gripper can be configured or replaced independently, so that it can be flexibly adjusted according to the needs of different magnetic measurement equipment, thereby improving the system's adaptability and scalability.

[0026] (6) This system is applicable to the automatic sample loading process of magnetic samples in geological and geophysical experiments and the magnetic testing of materials. By combining low magnetic interference mechanical structure, image recognition technology and adaptive control strategy, the automation level and operational stability of magnetic measurement process are improved.

[0027] Secondly, the present invention also provides a low-magnetic-interference robotic arm control method based on image recognition, the method comprising the following steps: acquiring an image containing a magnetic sample; recognizing the image using an image recognition module to determine the spatial pose parameters of the magnetic sample and the category identification information of the magnetic sample, wherein the spatial pose parameters include coordinates and rotation angles; generating control commands based on the spatial pose parameters and the category identification information of the magnetic sample, and controlling the robotic arm to execute the control commands.

[0028] It should be noted that this method possesses all the technical effects of the aforementioned control system, which will not be elaborated upon here. Attached Figure Description

[0029] The preferred embodiments of the present invention are described below with reference to the accompanying drawings, in which:

[0030] Figure 1 This is a schematic diagram of the gripper structure according to an embodiment of the present invention;

[0031] Figure 2 This is a flowchart illustrating how an image recognition module determines the spatial pose parameters and category identification information of a magnetic sample according to an embodiment of the present invention.

[0032] Figure 3 This is a flowchart illustrating an embodiment of the present invention that utilizes a channel attention submodule to adaptively enhance the spatial dimensions of each channel of a multi-scale feature map.

[0033] Figure label:

[0034] 1. Electric actuator; 2. End effector; 3. Gripper; 4. Shielding cover. Detailed Implementation

[0035] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0036] It should be noted that in the description of this invention, terms such as "upper" and "lower," indicating directional or positional relationships, are based on the directional or positional relationships shown in the accompanying drawings. This is merely for ease of description and does not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention. It should also be noted that in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0037] Currently, robotic arms used in industrial automation systems for magnetic measurement typically employ static path planning, which is prone to placement deviations, affecting measurement accuracy. In magnetic measurement scenarios, due to the unique characteristics of the samples (e.g., small size, smooth surface, reflective material), complex laboratory backgrounds, or uneven lighting, commonly used image recognition algorithms (such as YOLOv3 / v5, ResNet, etc.) are susceptible to problems like recognition drift, false detections, and inability to extract posture information. The running path determined based on the output sample position is often inaccurate, leading to robotic arm grasping failures or placement deviations. Furthermore, existing algorithms often focus only on position detection, lacking the ability to extract geometric parameters such as rotation angles and tilt directions, failing to meet the requirements for high-precision placement. Therefore, this invention uses an image recognition module to determine the spatial pose parameters and category identification information of the magnetic sample, and generates control commands based on these parameters and information. This enables precise positioning of the magnetic sample and determination of its posture, achieving high-precision grasping and placement.

[0038] In existing technologies, robotic arms used for sample grasping and delivery typically employ permanent magnet synchronous motors or ordinary servo systems, which inherently possess a certain magnetic field strength. In magnetic measurement scenarios, especially in ultra-sensitive systems such as SQUID, even weak magnetic interference can cause data drift or errors, affecting measurement results.

[0039] In one possible implementation, the present invention provides a low-magnetic-interference robotic arm control system based on image recognition. The system includes a robotic arm employing a six-degree-of-freedom serial structure to achieve a larger workspace and flexible posture adjustment. In applications with limited space or requiring only planar grasping and vertical movement, the robotic arm can also employ a three-axis parallel Delta structure. The robotic arm includes a robotic arm body and an end effector located at the end of the robotic arm body. The end effector includes a gripper. The arm body, transmission joints, and gripper of the robotic arm body are all made of non-magnetic materials, including but not limited to non-magnetic aluminum alloys, engineering plastics, or ceramic materials. The gripper is a two-finger parallel electric gripper with adjustable gripping force and supports angle adjustment and gripper length extension / retraction functions. The gripper spacing can be finely adjusted according to the size of the magnetic sample to adapt to magnetic samples of different sizes and depths. A Hall sensor is embedded in the end of the gripper for detecting and providing feedback on the gripping status.

[0040] The end effector uses a coreless brushless DC motor as the drive mechanism to provide the opening and closing driving force for the gripper. The coreless brushless DC motor has no iron core or is designed with a weak magnetic field. The transmission joint includes the joint structure and the joint motor used to drive the joint movement. The joint motor uses a low-magnetic coreless brushless DC motor. By combining the coreless brushless DC motor with a non-magnetic gear reduction system, the normal operation of the transmission joint is ensured, while the magnetic field generated by each motor during operation is reduced from the source. This reduces residual magnetism and transient electromagnetic disturbances during operation, solving the problem of magnetic field interference to magnetic measurement equipment caused by traditional robotic arms. Preferably, the coreless brushless motor can be positioned away from the detection area of ​​the magnetic measurement equipment, and the driving force is then transmitted to the gripper through a non-magnetic linkage or transmission mechanism to reduce the direct impact of the magnetic source on the magnetic measurement environment. Of course, the drive mechanism and joint motor of the end effector can also be other low-magnetic motors or mechanisms, such as pneumatic drive mechanisms, piezoelectric drive mechanisms, ultrasonic motors, low-residual magnet servo motors, low-residual magnet stepper motors, or low-residual magnet permanent magnet synchronous motors.

[0041] like Figure 1As shown, the robotic arm is also equipped with a shield 4 to shield and guide the magnetic field generated during the operation of the robotic arm's drive components, suppressing magnetic flux leakage and thus reducing the interference of the robotic arm on the magnetic measurement environment. Preferably, the shield is located on the outside of the end effector 2, which reduces the magnetic interference generated by the end effector 2 during operation. Of course, the shield can also be located on the outside of at least one of the electric actuators, joint motors, or transmission joints to shield the magnetic field generated by these drive components. The shield 4 is made of a high-permeability soft magnetic material, such as silicon steel, permalloy, amorphous or nanocrystalline alloys, iron-cobalt alloys, or ferrite soft magnetic materials such as manganese-zinc ferrite or nickel-zinc ferrite. Preferably, the shield is a double-layer structure, especially the shield located on the outside of the hollow cup brushless DC motor. The double-layer structure includes an inner shield shell and an outer shield shell, with a gap between the two shield shells to absorb and guide the leakage magnetic field generated during motor operation, thereby better preventing magnetic flux leakage. Obviously, the shielding cover can also be configured with three, four, or more layers.

[0042] In one possible implementation, the control signal lines of the robotic arm are all connected to the drive circuit via opto-isolation modules, and are electrically isolated and electromagnetically decoupled from the power supply circuit. The robotic arm uses a 24VDC industrial power supply with overcurrent and reverse connection protection. Furthermore, the motor power supply lines and signal lines are laid out in parallel loops, with wires in opposite current directions arranged adjacently to reduce the leakage magnetic field generated by the wires. Simultaneously, the drive motor and main electromagnetic components of the robotic arm are arranged symmetrically to reduce magnetic field coupling effects. Preferably, the control signal lines are connected to the control module of the control system (see details below) via opto-isolation modules. The control signal output from the control module is input to the input-side drive unit of the opto-isolation module. The input-side drive unit drives the light-emitting end (e.g., a light-emitting diode) of the optocoupler to emit light. The light signal is transmitted through an insulating isolation medium to the light-receiving end (e.g., a photodiode). The output-side signal recovery unit recovers the received light signal into the corresponding electrical control signal and outputs it to the drive circuit, achieving electrical isolation between the control module and the drive circuit, thereby reducing the impact of electromagnetic interference generated during motor operation on the control system. In addition, a shielding sleeve made of high permeability material can be installed on the outside of the cable path of the motor power supply line and control signal line. The shielding sleeve can shield and guide the magnetic field generated by the cable current, so as to reduce the influence of the magnetic field generated by the conductor on the magnetic measurement equipment.

[0043] By using non-magnetic materials to fabricate the robotic arm body, transmission joints, and grippers 3, employing a hollow cup brushless DC motor as the end effector 2, and installing shielding covers 4 on the outside of the end effector 2 and transmission joints, the magnetic field interference to the magnetic measurement equipment can be significantly reduced. Compared with traditional robotic arm solutions, magnetic interference can be reduced by 91.3%. Furthermore, experimental verification shows that at a typical safe distance of 30cm, the magnetometer is expected to avoid significant drift during long-term operation, with measurement data variability <0.05%. Over 400 hours of continuous testing on SQUID and Kappabridge equipment demonstrated stable and reliable interference suppression.

[0044] It should be noted that the drive mechanism of the end effector can also be a low-residual-magnetic-field permanent magnet synchronous motor. This type of motor is a specially designed permanent magnet synchronous motor that reduces the residual magnetic field strength in the rotor to a very low level after power failure through a unique magnetic circuit and structural design. Of course, the end effector can also use a pneumatic drive mechanism, such as a cylinder, pneumatic motor, or pneumatic servo system, especially in extreme magnetic field environments. Obviously, the end effector can also be a regular permanent magnet synchronous motor or a regular servo system. In this case, magnetic flux leakage can be prevented by setting multiple layers of shielding or magnetically conductive structures on the outside of the end effector, or by placing the end effector far away from the measurement area of ​​the magnetic measurement equipment and then transmitting the driving force to the end effector through a non-magnetic linkage or transmission mechanism, thus avoiding the magnetic field it generates from affecting the magnetic measurement environment.

[0045] In one possible implementation, the low-magnetic-interference robotic arm control system based on image recognition includes an image acquisition module, an image recognition module, and a control module. The image acquisition module acquires images of magnetic samples, and the image recognition module identifies the images and determines the spatial pose parameters and category identification information of the magnetic samples. The spatial pose parameters include coordinates (X, Y, Z) and rotation angle θ. Based on these spatial pose parameters, the magnetic samples can be accurately located and their poses can be accurately represented. After receiving the spatial pose parameters, the control module calculates the joint angle sequence of the robotic arm using an inverse kinematics algorithm based on the spatial pose parameters and the entry position of the magnetic measurement device. Simultaneously, it determines the corresponding magnetic sample based on the category identification information, generates control commands, and controls the robotic arm to execute the control commands. This enables precise grasping and delivery of magnetic samples, effectively improving the delivery accuracy.

[0046] Furthermore, the control module can also establish a magnetic field spatial distribution model by combining historically collected environmental magnetic field distribution data. It can then predict the magnetic field strength at each sampling point along the previously determined control command's operating path using the spatial interpolation method. This prediction can employ methods such as linear interpolation, inverse distance weighted interpolation, or three-dimensional interpolation to assess potential magnetic field changes along the path in advance. Next, the control module replans the path to avoid high-magnetic-field regions based on algorithms such as A* (A-star Algorithm) or RPT (Rapidly-exploring Random Tree), obtaining a corrected path that satisfies magnetic field constraints. This generates new control commands, which the robot arm then executes.

[0047] In one possible implementation, the image acquisition module can be an RGB camera, a depth camera, or a stereo camera. When the image acquisition module is a depth camera or a stereo camera, the acquired image contains depth information. The image recognition module can directly obtain the spatial pose parameters of the magnetic sample based on the depth information, thereby achieving three-dimensional localization and pose estimation of the magnetic sample. When the image acquisition module is an RGB camera, the spatial pose parameters of the magnetic sample can be obtained by combining visual markers or image feature-based recognition methods. Specifically, visual markers can be set on the magnetic sample or its carrier as a reference. The image recognition module can detect the visual markers and solve the spatial pose parameters of the magnetic sample based on the spatial correspondence of the marker points. Alternatively, the shape features, edge features, or texture features of the magnetic sample in the image can be extracted, and its spatial pose parameters can be directly determined by image recognition algorithms (such as edge detection algorithms, contour matching algorithms, template matching algorithms, pose estimation algorithms based on key point detection, or object detection and pose estimation algorithms based on deep learning). Through the above methods, even when using only an RGB camera, it is still possible to effectively acquire the spatial pose parameters of the magnetic sample, meeting the accuracy requirements of robotic arm grasping and delivery.

[0048] In one possible implementation, the image recognition module includes a reflection robust preprocessing submodule, a feature extraction submodule, a channel attention submodule, and a multi-task output submodule connected in sequence. The image of the magnetic sample is processed sequentially by these submodules, outputting the spatial pose parameters and category identification information of the magnetic sample. The spatial pose parameters include the magnetic sample's position coordinates (X, Y, Z) and rotation angle θ. The control module calculates the grasping angle, grasping position, grasping height, and placement depth of the end effector into the magnetic measurement device based on the position coordinates (X, Y, Z) and rotation angle θ. The category identification information includes the magnetic sample's number. During automated sample loading, multiple samples are typically numbered to form a task queue. The control module uses this number to correlate the magnetic sample with the samples in the task queue and can record the measurement results with the corresponding sample after measurement, establishing a "sample-measurement data" correspondence for sample task management, experimental data recording, and result traceability.

[0049] The reflection robust preprocessing submodule performs edge enhancement and reflection suppression on the input image. This is typically achieved by limiting the gradient magnitude of pixels corresponding to bright areas in the image, thus reducing false detections caused by metallic reflections. Specifically, the reflection robust preprocessing submodule first performs brightness analysis on the input image, identifying bright areas with brightness exceeding a preset threshold and marking these areas as reflection candidate regions. Subsequently, it limits the gradient magnitude of the pixels corresponding to the reflection candidate regions, ensuring that their gradient response does not exceed a preset upper limit. This prevents specular reflections from the magnetic sample surface from being misidentified as true boundaries during subsequent edge extraction. The limiting process can employ at least one of gradient truncation, gradient compression, or gradient decay based on a weighting function. After limiting the reflection candidate regions, the edge response of bright areas is weakened, while the edge information of the true contour region of the magnetic sample is preserved, thereby improving the stability of subsequent feature extraction and pose recognition.

[0050] The feature extraction submodule is used to extract multi-scale features from the preprocessed image. Through multi-scale feature fusion, it can comprehensively utilize information from different levels such as edges, textures, and shapes. Specifically, the feature extraction submodule employs a MagNet neural network, which includes a convolutional backbone and a multi-scale feature fusion structure. The convolutional backbone consists of multiple convolutional modules, at least one of which uses depthwise separable convolution. Furthermore, the number of parameters and computational cost of the convolutional backbone are less than a pre-set threshold based on the deployment requirements of edge computing devices. This allows for the preservation of detail perception capabilities for small-sized samples while reducing the number of parameters and computational cost, and improves its deployment adaptability on edge computing devices.

[0051] In practical applications, edge computing devices (such as Raspberry Pi) are often used in magnetic detection environments due to their low magnetic interference. However, the applicant found that edge computing devices have limited computing performance, while general-purpose models typically have a large number of parameters and complex calculations, leading to poor performance on edge computing devices and difficulty in ensuring recognition accuracy and stability, resulting in limited reliability in magnetic measurements. To address this, the feature extraction submodule of this application adopts the MagNet neural network, an optimized lightweight neural network architecture that includes a convolutional backbone and a multi-scale feature fusion structure. At least one of the multiple convolutional modules in the convolutional backbone uses depthwise separable convolution, and the number of parameters and computational cost of the convolutional backbone is less than a preset threshold. This significantly reduces the computational cost and number of parameters, maintaining high accuracy while significantly reducing computational cost and memory usage. It is smaller, faster, and easily deployed on edge computing devices. The multi-scale feature fusion structure can simultaneously aggregate low-level detailed features such as edges and textures with high-level semantic features such as object parts and categories, resulting in more accurate identification and localization of magnetic samples, especially adept at handling complex scenarios with large size variations and many small-sized samples.

[0052] Specifically, the MagNet neural network comprises a convolutional backbone and a multi-scale feature fusion structure. The convolutional backbone consists of multiple cascaded convolutional modules used for layer-by-layer feature extraction from the input image. At least one convolutional module employs a depthwise separable convolutional structure to improve the extraction capability of fine-grained features while reducing the number of parameters and computational complexity. Within the convolutional backbone, each layer extracts feature information at different scales through progressive downsampling operations to obtain multi-level feature representations including edge, texture, and semantic information. The multi-scale feature fusion structure is used to fuse feature maps of different scales. Through upsampling, feature concatenation, or element-wise weighting, it fuses low-level detailed features with high-level semantic features to enhance the representation capability of small-sized and complex-shaped targets. Preferably, batch normalization layers and non-linear activation functions are introduced into the convolutional modules to improve network training stability and feature representation capability.

[0053] It should be noted that the feature extraction submodule can also employ other possible algorithms, such as the lightweight VisionTransformer structure. Of course, in low-precision scenarios (where the requirement for accurate targeting is not so high), a combination of edge detection and template matching algorithms can also be used.

[0054] The channel attention submodule is used to adaptively enhance the spatial dimensions of each channel of the multi-scale feature map, for example, to achieve adaptive enhancement of the channel dimensions related to the edge and shape of the magnetic sample.

[0055] The multi-task output submodule is used to simultaneously output the spatial pose parameters and category identification information of magnetic samples based on enhanced multi-scale features. Specifically, the multi-task output submodule includes a shared feature input layer and multiple parallel output branches, including a pose parameter prediction branch and a classification branch, used for spatial pose parameter prediction and category identification, respectively. Specifically, the enhanced multi-scale features are input into the shared feature input layer for feature integration, and then passed to the pose prediction branch and the classification branch respectively. The pose prediction branch outputs the coordinates and rotation angle of the magnetic sample through regression, while the classification branch outputs the category identification information of the magnetic sample through classification, thus achieving synchronous output of multiple tasks. During training, the multi-task output submodule is optimized using a joint loss function, which includes position loss, pose loss, and classification loss. The position loss measures the deviation between the predicted position and the true position, the pose loss measures the rotation angle error, and the classification loss measures the difference between the predicted category and the true category. By weighting and combining these loss terms, collaborative optimization of different tasks is achieved. The weight coefficients of each loss term can be adjusted according to the importance of different tasks or the convergence during training. For example, the weights can be dynamically adjusted according to the changes in the loss values ​​of each task, so that each task remains relatively balanced during training, thereby improving the overall recognition accuracy and model stability.

[0056] like Figure 2 As shown in the figure, the workflow of the image recognition module is as follows: the image acquired by the input image acquisition module is processed sequentially through the reflection robust preprocessing submodule, the feature extraction submodule, the channel attention submodule, and the multi-task output submodule. The multi-task output submodule outputs the spatial pose parameters of the magnetic sample and the category identification information of the magnetic sample.

[0057] When deployed on Jetson-type edge computing devices, this image recognition module can output the spatial pose parameters of magnetic samples while maintaining the frame rate, enabling rapid and accurate identification of magnetic samples. Simultaneously, this image recognition module supports operation on the edge device itself and possesses both hot model updates and offline recognition capabilities.

[0058] In one possible implementation, the channel attention submodule includes a global average pooling subunit, a weight generation subunit, and a channel weighting subunit. The global average pooling subunit performs spatial average pooling on the multi-scale feature map output by the feature extraction submodule to generate channel description vectors. Specifically, the global average pooling subunit performs spatial average pooling on each channel of the multi-scale feature map, compressing the features of each channel into a scalar to generate a channel description vector. The weight generation subunit includes a sequentially connected activation function layer, a fully connected layer, and a sigmoid function layer, used to perform nonlinear mapping on the channel description vectors and generate corresponding channel weight vectors. The activation function layer can use ReLU, Gaussian error linear unit (GELU), or other nonlinear activation functions to enhance feature representation. The channel weighting subunit performs channel-level multiplication on the input multi-scale feature map and the channel weight vector, multiplying each channel of the multi-scale feature map by the corresponding weight value in the channel weight vector to achieve adaptive weighting of the features of each channel, resulting in a refined feature map.

[0059] like Figure 3 As shown, the channel attention submodule performs adaptive spatial dimension enhancement on each channel of the multi-scale feature map, specifically including:

[0060] Input multi-scale feature map: Input the multi-scale feature map extracted by the feature extraction submodule into the channel attention submodule;

[0061] Activation function layer: Performs nonlinear transformation on the channel description vector to enhance feature representation ability;

[0062] Global average pooling subunit: Performs global average pooling on each channel of the multi-scale feature map in the spatial dimension, compressing the features of each channel into a scalar and generating a channel description vector;

[0063] ReLU activation function layer: performs non-linear transformation on features to enhance their expressive power;

[0064] Fully connected layer: Performs feature combination and compression between channels;

[0065] Sigmoid activation function layer: outputs channel weight vectors, which are between 0 and 1, forming a channel attention mask;

[0066] Channel weighted sub-unit: Each channel of the input multi-scale feature map is multiplied by the weight value at the corresponding position of the channel weight vector to weight each channel, and the refined feature map is output.

[0067] Continue to refer to Figure 1The end effector of the robotic arm also includes an electric push rod 1. One end of the electric push rod 1 is connected to the end of the robotic arm, and the other end is fixedly connected to the gripper 3. The electric push rod 1 extends and retracts along its length under the action of a hollow cup brushless DC motor. This extension and retraction of the electric push rod 1 drives the gripper 3 to move along its length, allowing the gripper 3 to approach or move away from the detection area of ​​the magnetic measurement equipment. This adapts to the magnetic sample grasping needs under different spatial positions and magnetic field environments, ensuring grasping accuracy while reducing the impact of the drive components on the magnetic measurement environment. A magnetic field sensor is installed near the gripper 3 on the robotic arm body to detect the ambient magnetic field strength at the location of the gripper 3 in real time. This magnetic field sensor can be a Hall sensor, a fluxgate sensor, or other magnetic field detection device. By arranging the magnetic field sensor near the gripper 3, the magnetic field environment at the magnetic sample grasping position can be more accurately reflected, providing a basis for subsequent gripper 3 extension / retraction adjustment and path optimization. It should be noted that the electric push rod 1 and the gripper 3 are connected by a guide connector. This guide connector is used to connect the electric push rod 1 and the gripper 3, transmit the extension and retraction driving force, and guide the movement of the gripper, so as to prevent the gripper 3 from swaying or shaking during the extension and retraction process.

[0068] In one possible implementation, the control system further includes an adjustment module communicatively connected to the control module. The adjustment module includes a first adjustment submodule and a second adjustment submodule. The first and second adjustment submodules adjust the extension length and opening angle of the grippers based on comparisons of the magnetic field strength with a first threshold and a second threshold, respectively, and the size of the magnetic sample. The first threshold B1 is the upper limit of the safe magnetic field, and the second threshold B2 is the upper limit of the magnetic field of the detection cavity of the magnetic measuring device. In this embodiment, the detection cavity refers to the detection space inside the magnetic measuring device used to place the sample to be tested.

[0069] Specifically, the first adjustment submodule is used to adjust the extension length of the gripper based on the magnetic field strength B. If the magnetic field strength B is less than or equal to a first threshold B1, the extension length of the gripper is adjusted to the minimum retraction length L. min This is to improve structural rigidity. It should be noted that the extension length of the gripper refers to the distance in the extension direction between the end of the gripper and its connection point with the end of the robotic arm. This distance is determined by the extension stroke of the electric actuator and is used to characterize the extent to which the gripper extends in space.

[0070] If the magnetic field strength B is greater than the first threshold B1 and less than or equal to the second threshold B2, the extension length of the gripper is adjusted according to a preset function, where the preset function is a linear function or a piecewise function L=f(B), to keep the gripper tip away from the main interference source. For example, it can be set according to the following linear relationship: when B increases from B1 to B2, the gripper length increases from L... min Linear transformation to L max That is, L=Lmin +(B-B1) / (B2-B1)·(L max -L min Under the above linear relationship, when the magnetic field strength B takes any intermediate value between the first threshold B1 and the second threshold B2, the extension length of the gripper can also be directly calculated using this linear relationship. For example, when the magnetic field strength is B3 and B1 < B3 < B2, the corresponding gripper extension length is: L = Lmin + (B3 - B1) / (B2 - B1) · (L max -L min This allows for continuous adjustment of the gripper extension length as the magnetic field strength changes. Therefore, when the magnetic field strength is in different ranges, the gripper extension length can be dynamically calculated based on real-time measurements, thus achieving continuously adjustable extension and retraction control.

[0071] If the magnetic field strength B is greater than the second threshold B2, the comparison result is fed back to the control module, the point is marked as a high magnetic field region, and then the control module replans the running path.

[0072] It should be noted that IMU and visual servoing algorithms can be combined when performing path correction.

[0073] The second adjustment submodule is used to adjust the opening angle of the grippers based on the size of the magnetic sample and the magnetic field strength. The size W of the magnetic sample is obtained by the image recognition module based on the acquired image, for example, by determining its characteristic dimensions through contour extraction, boundary detection, or circumscribed rectangle fitting of the target sample, and then converting it into the actual size W by combining the camera calibration parameters. Specifically, there is a preset mapping relationship α between the opening angle α of the grippers and the size W of the magnetic sample and the magnetic field strength B. When the size W of the magnetic sample is large, the opening angle α of the grippers is appropriately increased to ensure the clamping space. When the magnetic field strength B is large, the opening angle α of the grippers is appropriately decreased to shorten the force arm, improve clamping stability, and reduce the influence of magnetic interference. When the magnetic field strength B increases, the opening angle α of the grippers is adjusted from α' to α'' in preset steps. max Decrease step by step to α min Preferably, the preset step size can be α. max 5% to 10%, that is, whenever the magnetic field strength increases to a new range, α is reduced from the current angle. max ×5% or α max ×10%. Of course, the preset step size can also be a fixed angle value, such as decreasing by 2° or 5° each time, to adapt to gripper designs with different structural dimensions. Through the above adjustment method, the gripper opening angle can be dynamically adjusted according to the change of magnetic field strength, thereby reducing magnetic field interference while ensuring gripping reliability.

[0074] This approach links the adjustment of the gripper's extension length and opening angle with the control module's path planning, forming a closed-loop control strategy of "magnetic field sensing - threshold judgment - path correction / length / opening angle adjustment - gripping execution." When the ambient magnetic field strength is detected to be close to or exceed a preset threshold, the robotic arm's path is adjusted to avoid high magnetic field areas. The extension length of the gripper is changed to move the drive motor away from the magnetic measurement equipment's detection area. Simultaneously, the opening angle of the gripper is adjusted to reduce current fluctuations caused by changes in motor load, thereby reducing the magnetic field strength generated during the robotic arm's operation. Without adding additional shielding structures, this achieves active suppression of magnetic interference, reducing the robotic arm's disturbance to the magnetic measurement environment.

[0075] In one possible implementation, a host computer is also included, which can provide a unified operation and data management platform, including: (1) a graphical interface for operation control, task scheduling and module monitoring; (2) multi-protocol communication: supporting communication with robotic arms, motors and magnetic measuring devices via serial port, CAN or GPIO; (3) information visualization: real-time display of recognition results, motion status and task logs; (4) data recording and tracking: automatically recording captured images, sample information and timestamps, supporting subsequent tracking and analysis.

[0076] In summary, this application combines an image recognition module optimized for magnetic sample environments with an adaptive control strategy that adjusts the extension and opening angles of the grippers based on magnetic field strength feedback. This ensures recognition accuracy while optimizing the materials and structure of the robotic arm, reducing the risk of magnetic field coupling between the robotic arm and the magnetic measurement device. As a result, it can maintain a high grasping success rate and high measurement stability under complex magnetic field and lighting conditions.

[0077] In one possible implementation, the present invention also provides a low-magnetic-interference robotic arm control method based on image recognition, characterized in that the method includes the following steps:

[0078] S1: Acquire an image containing the magnetic sample;

[0079] S2: Use the image recognition module to recognize the image and determine the spatial pose parameters and category identification information of the magnetic sample. The spatial pose parameters include coordinates and rotation angles.

[0080] S3: Generate control commands based on spatial pose parameters and category identification information, and control the robotic arm to execute the control commands.

[0081] This method is applicable to the aforementioned low-magnetic-interference robotic arm control system based on image recognition.

[0082] As one specific implementation, the low-magnetic-interference robotic arm control system based on image recognition includes an image acquisition unit, an image recognition unit, a robotic arm, and a control circuit unit. The image acquisition unit can be an industrial camera, such as a 12-megapixel RGB camera, used to acquire images of the sample tray containing magnetic samples. The image recognition unit, located in a Jetson Xavier NX edge computing device or a host computer, is used to process the images acquired by the industrial camera to obtain the spatial pose parameters and category identification information of the magnetic samples. The control circuit unit, located in the robotic arm, includes a control module, a driver board, and a communication interface. It receives the recognition results output by the image processing unit and controls the robotic arm to perform corresponding actions. The control circuit unit communicates with the edge computing device or host computer through the communication interface.

[0083] Taking the automated grasping and placement of magnetic samples from a sample tray to the detection chamber of a magnetic measurement device using a robotic arm as an example, firstly, an industrial camera acquires images of the sample tray and transmits the acquired images to an image recognition unit. The image recognition unit identifies the spatial pose parameters and category identification information of the target sample and sends the recognition results to the control circuit unit. The control module in the control circuit unit calculates the grasping path of the robotic arm based on the spatial pose parameters and category identification information, and controls the movement of the robotic arm via a drive board, causing the end effector to move above the target sample. Subsequently, the control module controls the drive mechanism of the end effector, which in turn controls the gripper to open and move downwards to the grasping position of the target sample, achieving sample grasping by closing the gripper. During the grasping process, a magnetic field sensor monitors the ambient magnetic field strength near the end of the robotic arm in real time. When the detected magnetic field strength exceeds a preset threshold, the control module adjusts the robotic arm's running path according to a pre-set magnetic field interference suppression control strategy to avoid areas with high magnetic field strength. After path adjustment, if the magnetic field strength exceeds the preset threshold again during movement or at the target location, the control module sends a control command to the adjustment module. The adjustment module adaptively adjusts the gripper's posture parameters based on the detected magnetic field strength to further reduce the robotic arm's interference with the magnetic measurement environment. After grasping, the robotic arm moves the target sample to the entrance of the detection chamber of the magnetic measurement equipment along the planned path and places the sample in the designated position, thus completing an automatic sample loading operation. Alternatively, if the magnetic field strength exceeds the preset threshold, the adjustment module can first adaptively adjust the gripper's posture parameters. If the magnetic field strength still exceeds the preset threshold after adjustment, the control module adjusts the robotic arm's running path according to a pre-set magnetic field interference suppression control strategy. Obviously, when the magnetic field strength is detected to exceed the preset threshold, the operating path of the robotic arm can be adjusted by the control module and the attitude parameters of the gripper can be adjusted by the adjustment module.

[0084] The robotic arm employs a six-degree-of-freedom serial structure. The arm body is composed of non-magnetic aluminum alloy and polycarbonate, and the grippers are two-finger parallel electric grippers with embedded Hall effect sensors at their ends for detecting and providing feedback on the gripping status. The end effector's drive mechanism, a hollow cup brushless DC motor, is positioned away from the magnetic measurement equipment. A flexible connecting rod, made of non-magnetic material, transmits power from the drive mechanism to the grippers. This reduces the influence of the magnetic field generated by the drive mechanism on the magnetic measurement equipment, allowing the end effector to perform magnetic sample gripping operations even when positioned close to the detection cavity of the magnetic measurement equipment.

[0085] The end effector's drive mechanism and all joint motors use coreless brushless DC motors combined with a non-magnetic gear reduction system. The drive mechanism and shutdown motor of the end effector are equipped with double-layer shielding. All signal lines are optically isolated and electromagnetically decoupled from the power supply system.

[0086] The control circuit unit includes: (1) Control module: based on STM32MCU, used for motion control and communication coordination; (2) Driver board: supports six-channel BLDC motor control with encoder feedback; (3) Communication interface: connected to the driver module via CAN bus, communicates with the magnetic measurement equipment via RS-232 or GPIO, and communicates with the host computer via USB interface; (4) Safety module: set status indicator lights, limit protection switches and emergency stop buttons to ensure safe operation.

[0087] The power supply system uses a 24VDC industrial power supply to provide stable power to all electrical equipment in the system, and has overcurrent protection and reverse connection protection functions. The industrial camera connects to the host computer via a GigE interface to transmit acquired image data to the host computer for processing. The image recognition module is deployed in the host computer to recognize the image data and output the spatial pose parameters and category identification information of the magnetic sample. The control circuit unit communicates with the host computer via a USB interface to receive image recognition results and control commands; the control circuit unit also connects to the robotic arm driver via a CAN bus to control the movement of each joint of the robotic arm; simultaneously, the control module and driver board communicate with the magnetic measurement equipment via an RS-232 serial port to send measurement commands and exchange status information.

[0088] Based on the above system, the process of automatically loading samples onto a magnetometer using a robotic arm and completing magnetic detection includes the following steps:

[0089] S1, System Initialization: The host computer control system completes communication connections and status checks with the image recognition unit, control circuit unit, and magnetometer. After system initialization, the robotic arm maintains its default posture and remains in standby mode, the image recognition unit loads, and all modules enter a linked standby state.

[0090] S2, Image Acquisition and Recognition:

[0091] S21 uses an industrial camera to capture RGB images of the tray area and sends the images to an edge computing device;

[0092] S22, the edge computing device first performs normalized edge enhancement and highlight area suppression on the image through the reflection robust preprocessing submodule to obtain the intermediate image after reflection suppression;

[0093] S23: Input the intermediate image into the feature extraction submodule and the channel attention submodule to obtain a multi-scale feature map containing sample edge, shape and texture information;

[0094] S24: Input the multi-scale feature map into the multi-task output submodule to regress the spatial pose parameters and category recognition information of the magnetic sample, including three-dimensional coordinates (X,Y,Z), rotation angle θ, and sample ID.

[0095] S25: Based on the weighted results of position loss, pose loss and classification loss in the loss function, train or update the multi-task output submodule online, and write the recognition results into the task queue in sequence.

[0096] S3, Path Planning:

[0097] S31: The control module uses an inverse kinematics algorithm to solve the joint angle sequence of the robotic arm based on the spatial pose parameters of the magnetic sample, the category identification information, and the entry position of the magnetometer, and obtains the initial running path.

[0098] S32: Based on historically collected environmental magnetic field distribution data, predict the magnetic field strength of each sampling point on the initial trajectory;

[0099] S33: The path planning module replans the trajectory to avoid the high magnetic field region based on the RRT algorithm, and obtains a corrected path that meets the magnetic field constraints.

[0100] S4, Sample gripping and adaptive magnetic field adjustment:

[0101] S41: Control the end of the robotic arm to move along the corrected path to above the target sample, and collect the current magnetic induction intensity B in real time through the triaxial magnetic field sensor set at the end of the robotic arm body;

[0102] S42: The control module determines the gripper extension length L and opening angle α based on the relationship between B and the thresholds B1 and B2, and drives the electric push rod and motor to adjust the gripper posture.

[0103] S43: The gripper performs the gripping action in the sequence of "opening → pressing down → closing → lifting", and at the same time monitors the gripping status of the gripper through the Hall sensor. If the gripping fails, the gripping process is triggered again.

[0104] S44: After the sample is lifted, the robotic arm carries the sample along the corrected path to the entrance of the magnetometer, collects the magnetic field strength B again and fine-tunes L and α, so that the sample is slowly sent into the detection cavity and released under low magnetic interference.

[0105] Step S5, Magnetic Measurement Start:

[0106] After clamping is complete, the control circuit unit sends a start command to the magnetometer via GPIO or serial interface to trigger the measurement process. The host computer records the sample number, start time, and status feedback, and synchronously writes the data to the database.

[0107] Step S6, task loop check:

[0108] The host computer determines whether there are still unprocessed samples in the task queue. If so, it automatically returns to step S2 to continue the next round of identification and capture process.

[0109] Finally, based on the task queue status and magnetometer feedback information, the automatic identification, adaptive grasping, and magnetic measurement of all samples are completed, realizing a high-precision batch processing workflow for weakly magnetic samples.

[0110] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such order. They can be executed simultaneously (in parallel) or in reverse order. These simple changes are all within the protection scope of this application.

[0111] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A low-magnetic-interference robotic arm control system based on image recognition, characterized in that, include: The image acquisition module is used to acquire images of magnetic samples; An image recognition module is used to recognize the image acquired by the image acquisition module, determine the spatial pose parameters of the magnetic sample and the category identification information of the magnetic sample, wherein the spatial pose parameters include coordinates and rotation angles; The control module is used to receive the spatial pose parameters and the category identification information, generate control commands based on the spatial pose parameters and the category identification information, and control the robotic arm to execute the control commands.

2. The low-magnetic-interference robotic arm control system based on image recognition as described in claim 1, characterized in that, The image recognition module includes a reflection-robust preprocessing submodule, a feature extraction submodule, a channel attention submodule, and a multi-task output submodule connected in sequence. The reflection-robust preprocessing submodule performs edge enhancement and reflection suppression on the input image. The feature extraction submodule extracts multi-scale feature maps from the preprocessed image. The feature extraction submodule uses a MagNet neural network, which includes a convolutional backbone and a multi-scale feature fusion structure. The convolutional backbone consists of multiple convolutional modules, at least one of which employs depthwise separable convolution. The channel attention submodule is used to adaptively enhance the spatial dimension of each channel of the multi-scale feature map, and the multi-task output submodule is used to synchronously output the spatial pose parameters and the category identification information based on the enhanced multi-scale feature map.

3. The low-magnetic-interference robotic arm control system based on image recognition as described in claim 2, characterized in that, The reflection robust preprocessing submodule is used to identify the bright areas in the input image and suppress the gradient magnitude of the pixels corresponding to the bright areas.

4. The control system according to claim 2, characterized in that, The channel attention submodule includes: a global average pooling subunit, used to perform global average pooling on each channel of the multi-scale feature map in the spatial dimension, compressing the features of each channel into a scalar to generate a channel description vector; a weight generation subunit, including an activation function layer, a fully connected layer and a sigmoid function layer connected in sequence, used to generate a corresponding channel weight vector based on the channel description vector; and a channel weighting subunit, used to perform channel-level multiplication on the input multi-scale feature map and the channel weight vector to weight each channel.

5. The low-magnetic-interference robotic arm control system based on image recognition according to any one of claims 1 to 4, characterized in that, The image acquisition module is an RGB camera, a depth camera, or a stereo camera.

6. The low-magnetic-interference robotic arm control system based on image recognition according to any one of claims 1 to 4, characterized in that, The arm body, drive joints, and / or grippers of the robotic arm are made of non-magnetic materials; and / or the drive mechanism of the end effector of the robotic arm and / or the joint motor of the drive joint are hollow cup brushless DC motors.

7. The low-magnetic-interference robotic arm control system based on image recognition according to any one of claims 1 to 4, characterized in that, The robotic arm is equipped with a shield, which is located at least on the outside of the end effector and / or drive joint of the robotic arm.

8. The low-magnetic-interference robotic arm control system based on image recognition according to any one of claims 1 to 4, characterized in that, The robotic arm includes a robotic arm body and an end effector disposed at the end of the robotic arm body. The end effector includes a gripper. A magnetic field sensor is disposed on the robotic arm body near the gripper. The magnetic field sensor is used to collect the ambient magnetic field strength at the location of the gripper. The control system also includes an adjustment module, which is communicatively connected to the control module. The adjustment module includes a first adjustment submodule and a second adjustment submodule. The first adjustment submodule is used to adjust the extension length of the gripper based on the magnetic field strength. The second adjustment submodule is used to adjust the opening angle of the gripper based on the size of the magnetic sample and the magnetic field strength.

9. The low-magnetic-interference robotic arm control system based on image recognition as described in claim 8, characterized in that, The first adjustment submodule is used to adjust the extension length of the gripper to the minimum retraction length when the magnetic field strength is less than or equal to a first threshold, and to adjust the extension length of the gripper according to a preset function when the magnetic field strength is greater than the first threshold and less than or equal to a second threshold. The preset function is a linear function or a piecewise function, the first threshold is the upper limit of the safe magnetic field, and the second threshold is the upper limit of the magnetic field of the detection cavity of the magnetic measuring device.

10. A low-magnetic-interference robotic arm control method based on image recognition, characterized in that, The method includes the following steps: acquiring an image containing a magnetic sample; recognizing the image using an image recognition module to determine the spatial pose parameters of the magnetic sample and the category identification information of the magnetic sample, wherein the spatial pose parameters include coordinates and rotation angles; generating control commands based on the spatial pose parameters and the category identification information, and controlling the robotic arm to execute the control commands.

Images