Software robot control method and device, software robot, and electronic device

Abstract

This invention relates to a control method, device, soft manipulator, and electronic device for a soft robotic arm. The method includes: receiving first data and a first image, wherein the first data is pressure sensor data and IMU sensor data corresponding to the manipulator, and the first image is image data measured by an image sensor inside the manipulator. The first data is processed to obtain second data, which includes load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data. The first image after edge detection is identified using a convolutional neural network to obtain volume distribution data at different locations inside the manipulator. A deep reinforcement learning algorithm is invoked to analyze and process the second data and volume distribution data to obtain first control parameters, which are PID control parameters used to control the air intake and exhaust volume of an air pump connected to the manipulator, achieving reliable and stable grasping of fragile and easily damaged objects.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, and in particular to a soft robotic arm control method, device, soft robotic arm, and electronic device. Background Technology

[0002] Robotic arms can mimic certain movements and functions of human hands and arms to grasp, move objects, or operate tools according to fixed programs. They are a new type of device developed in mechanized and automated production processes. In modern production, robotic arms are widely used in automated production lines. The research and production of robots has become a rapidly developing emerging technology in the high-tech field, promoting the development of robotic arms and enabling them to better integrate with mechanization and automation. Although robotic arms are not as dexterous as human hands, they possess the advantages of continuous repetitive work and labor, tirelessness, fearlessness of danger, and greater lifting strength than human hands. Therefore, robotic arms have received attention from many sectors and are being used more and more widely.

[0003] Currently, most robotic arms in use are rigid robotic arms, whose traditional functions and forms have been widely applied in human society's production. These rigid robotic arms are mainly made of metal materials and are often responsible for structural and repetitive tasks in industrial settings. However, modern robotic arms are developing towards higher positioning accuracy, higher flexibility, and higher human-machine interaction. Due to the excessive rigidity of the materials used in rigid robotic arms, it is difficult to reliably grasp some delicate or fragile items.

[0004] Therefore, traditional rigid robotic arms have difficulty achieving reliable and stable gripping of fragile and easily damaged objects. Summary of the Invention

[0005] Therefore, it is necessary to provide a control method, device, soft manipulator, and electronic device for a soft manipulator that can reliably and stably grasp fragile and easily damaged objects, in order to address the above-mentioned technical problems.

[0006] This invention provides a method for controlling a soft robotic arm, the method comprising:

[0007] Receive first data and first image, wherein the first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm;

[0008] The first data is processed to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0009] The first image after edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm.

[0010] The second data and volume distribution data are analyzed and processed using a deep reinforcement learning algorithm to obtain a first control parameter, which is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

[0011] In one embodiment, the process of processing the first data to obtain the second data further includes:

[0012] Based on the IMU sensor data, a first vibration signal is obtained, wherein the first vibration signal is the vibration signal of the object being grasped by the robotic arm;

[0013] Based on the first vibration signal, a loss function corresponding to the vibration error is obtained, and the loss function is used to describe the suppression effect on the first vibration effect.

[0014] In one embodiment, the method further includes:

[0015] Based on the first data and the first image, the first state of the grasped object is obtained, and the first state is the current state of the robotic arm;

[0016] When the object being grasped is in a first state and the value of the corresponding loss function exceeds a first threshold, the convolutional neural network is invoked to fit the first state to obtain a first behavior. The first behavior is the best behavior of the robotic arm when the loss function exceeds the first threshold in the first state.

[0017] The first action is used to increase the first control parameter.

[0018] In one embodiment, the method further includes:

[0019] When the object being grasped is in a first state and the value of the corresponding loss function is lower than a second threshold, the convolutional neural network is invoked to fit the first state to obtain a second behavior, which is the best behavior of the robotic arm when the loss function is lower than the second threshold in the first state.

[0020] The second action is used to reduce the first control parameter.

[0021] In one embodiment, the step of identifying the first image after edge detection processing using a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm further includes:

[0022] Acquire image data before and after the robotic arm's grasping motion;

[0023] Edge detection is performed on the image data before and after the grasping motion of the robotic arm to obtain the image to be identified after edge detection.

[0024] In one embodiment, the step of recognizing the first image after edge detection processing using a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm includes:

[0025] The convolutional neural network is invoked to identify the image to be recognized in order to obtain the volume distribution data of the robotic arm before and after the grasping motion.

[0026] In one embodiment, the step of invoking a deep reinforcement learning algorithm to analyze and process the second data and the volume distribution data includes:

[0027] The PID algorithm is invoked to calculate the pulse width modulation parameters by analyzing the results of the deep reinforcement learning algorithm.

[0028] The pulse width modulation parameters are sent to the lower-level computer connected to the robotic arm via a transmission control protocol to control the robotic arm to perform grasping movements.

[0029] The present invention also provides a soft robotic arm control device, the device comprising:

[0030] The data receiving module is used to receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm.

[0031] A first processing module is used to process the first data to obtain second data, wherein the second data is load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data;

[0032] The image recognition module is used to recognize the first image after image edge detection processing through a convolutional neural network in order to obtain volume distribution data at different locations inside the robot arm;

[0033] The second processing module is used to call a deep reinforcement learning algorithm to analyze and process the second data and volume distribution data to obtain a first control parameter. The first control parameter is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

[0034] The present invention also provides a soft robotic arm, controlled by any of the soft robotic arm control methods described above, comprising:

[0035] The robotic arm body consists of soft fingers, a support structure, and an air drive tube. The support structure is used to support and fix multiple soft fingers. The air drive tube is connected to the soft fingers and controls the extension and bending of the soft fingers by air intake and exhaust.

[0036] The lower-level machine is electrically connected to the robot body and is used to receive sensor data from the robot body and directly control the gripping motion of the robot body by controlling the air intake and exhaust of the pneumatic drive tube.

[0037] The host computer, electrically connected to the slave computer, is used to receive sensor data from the slave computer via a transmission control protocol and process the sensor data to obtain control parameters for controlling the gripping motion of the robotic arm body, and send the control parameters to the slave computer via the transmission control protocol.

[0038] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the soft robotic arm control method as described above.

[0039] The aforementioned soft manipulator control method, device, soft manipulator, and electronic equipment receive pressure sensor data, IMU sensor data, and image data from inside the soft manipulator. They then process the pressure sensor data and IMU sensor data to obtain load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data. Subsequently, a convolutional neural network is used to identify the image data processed by image edge detection inside the soft manipulator to obtain volume distribution data at different locations inside the manipulator. Finally, a deep reinforcement learning algorithm is invoked to comprehensively analyze and process the load distribution data, vibration error, and volume distribution data inside the soft manipulator to obtain PID control parameters for controlling the air output of the air pump connected to the soft manipulator, thereby controlling the grasping motion of the soft manipulator. This method combines a soft robotic arm with pressure sensors, image sensors, and IMU sensors to achieve comprehensive monitoring of the soft robotic arm's motion state. It also controls the robotic arm's gripping motion through deep reinforcement learning and PID control parameters, enabling the robotic arm to flexibly adjust its gripping motion in different states. This avoids drastic changes in gripping force and vibration during the gripping process, thus achieving reliable and stable gripping of fragile and easily damaged objects. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 This is one of the schematic diagrams of the soft robotic arm control method provided by the present invention;

[0042] Figure 2 The second schematic diagram of the soft robotic arm control method provided by the present invention;

[0043] Figure 3 The third schematic diagram of the soft robotic arm control method provided by the present invention;

[0044] Figure 4 The fourth schematic diagram of the soft robotic arm control method provided by the present invention;

[0045] Figure 5 Fifth schematic diagram of the soft robotic arm control method provided by the present invention;

[0046] Figure 6 This is a schematic diagram of the soft robotic arm control system structure in a specific embodiment of the present invention;

[0047] Figure 7 This is a schematic diagram of the soft robotic arm host computer control system structure in a specific embodiment of the present invention;

[0048] Figure 8 A schematic diagram of the soft robotic arm control device provided by the present invention;

[0049] Figure 9 An internal structural diagram of the computer device provided by the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0051] The following is combined Figures 1-9 The present invention describes a soft robotic arm control method, apparatus, soft robotic arm, and electronic device.

[0052] like Figure 1 As shown, in one embodiment, a soft robotic arm control method includes the following steps:

[0053] Step S110: Receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm.

[0054] Among them, the IMU (Inertial Measurement Unit) is mainly used as a sensor to detect and measure acceleration and rotational motion. Its principle is based on the law of inertia. It measures the acceleration in the three directions of movement (X, Y, and Z) through the accelerometer contained therein. The velocity is obtained through a single integration and the position is obtained through a double integration.

[0055] Specifically, the host computer receives pressure data measured by the pressure sensor on the soft robotic arm, motion data measured by the IMU sensor on the soft robotic arm, and image data of the inside of the robotic arm measured by the image sensor inside the soft robotic arm.

[0056] Step S120: Process the first data to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0057] Specifically, the host computer converts the pressure sensor data and IMU sensor data obtained in step S110 on the soft manipulator to obtain load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data.

[0058] The pressure sensor employs a mixed-site metal-carbon nanofilm pressure sensor, embedded at the edge of a soft robotic arm. It utilizes magnetron sputtering and ultraviolet lithography to deposit a 70nm thick titanium alloy electrode as the negative electrode on the left side and a 70nm thick titanium alloy electrode as the positive electrode on the right side of the stretchable mixed-site metal-carbon nanofilm sensor. The effective bending angle of the mixed-site metal-carbon nanofilm pressure sensor is 0° to 180°. Under the external force generated by the bending of the soft finger, the mixed-site metal-carbon nanofilm sensor undergoes strain, causing dislocations to bond between the metal nanoparticles and graphene in the carbon film. This reduces the conductive pathway, leading to an increase in sensor resistance. However, due to the good conductivity of the mixed-site metal-carbon nanoparticles and the increased conductivity of the conductive polyaniline particles, the conductive network can be maintained. The change in sensor resistance results in corresponding changes in the voltage signal and the load distribution at the corresponding location, thus reflecting the pressure distribution during the bending of the soft finger.

[0059] Step S130: The first image after image edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robot arm.

[0060] Specifically, the host computer first performs image edge detection on the image data inside the soft finger measured by the image sensor to determine the images corresponding to different parts of the soft manipulator. Then, it uses a convolutional neural network to identify the image data after edge detection to obtain the volume distribution data of different locations inside the soft manipulator.

[0061] It should be noted that during the movement of the soft robotic hand, the internal volume of the soft robotic hand is constantly changing due to the bending and straightening of the robotic hand's fingers, the intake and exhaust of air inside the robotic hand, and the grasping process of the soft robotic hand on the object being grasped.

[0062] Step S140: Call a deep reinforcement learning algorithm to analyze and process the second data and volume distribution data to obtain the first control parameter. The first control parameter is a PID control parameter used to control the air intake and output of the air pump connected to the robot arm.

[0063] Specifically, the host computer calls a deep reinforcement learning algorithm to comprehensively analyze and process the load distribution data corresponding to the pressure sensor data obtained in step S120, the vibration error corresponding to the IMU sensor data, and the volume distribution data corresponding to the soft manipulator obtained in step S130, and finally obtains the PID control parameters for controlling the air intake and exhaust volume of the air pump connected to the soft manipulator.

[0064] It should be noted that when the soft robotic arm vibrates after grasping the object, the host computer appropriately controls the air intake and output of the air pump connected to the soft robotic arm through corresponding PID control parameters, so as to ensure that the soft robotic arm's grasp of the object meets the requirements of stability and reliability, and reduces the vibration of the object.

[0065] The aforementioned soft manipulator control method receives pressure sensor data, IMU sensor data, and image data from inside the soft manipulator. It then processes the pressure sensor and IMU sensor data to obtain load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data. Subsequently, a convolutional neural network is used to identify the image data processed by edge detection inside the soft manipulator to obtain volume distribution data at different locations within the manipulator. Finally, a deep reinforcement learning algorithm is invoked to comprehensively analyze the load distribution data, vibration error, and volume distribution data within the soft manipulator, obtaining PID control parameters for controlling the air output of the air pump connected to the soft manipulator, thereby controlling the gripping motion of the soft manipulator. This method combines a soft robotic arm with pressure sensors, image sensors, and IMU sensors to achieve comprehensive monitoring of the soft robotic arm's motion state. It also controls the robotic arm's gripping motion through deep reinforcement learning and PID control parameters, enabling the robotic arm to flexibly adjust its gripping motion in different states. This avoids drastic changes in gripping force and vibration during the gripping process, thus achieving reliable and stable gripping of fragile and easily damaged objects.

[0066] like Figure 2 As shown, in one embodiment, the soft robotic arm control method provided by the present invention processes first data to obtain second data, and includes the following steps beforehand:

[0067] Step S210: Based on IMU sensor data, acquire a first vibration signal, which is the vibration signal of the object being grasped by the robotic arm.

[0068] Specifically, the host computer acquires the vibration signal of the grasped object based on the motion data of the soft robotic arm measured by the IMU sensor.

[0069] It should be noted that the motion data of the soft robotic arm measured by the IMU sensor is constantly changing. That is, the acceleration, velocity and position of the soft robotic arm are constantly changing. The frequency of the changes in the motion data of the host computer and the soft robotic arm can be used to obtain the vibration signal of the soft robotic arm, that is, the first vibration signal.

[0070] Step S220: Based on the first vibration signal, obtain the loss function corresponding to the vibration error. The loss function is used to describe the suppression effect on the first vibration signal.

[0071] Specifically, based on the first vibration signal obtained in step S210, the host computer obtains the loss function corresponding to the vibration error. This loss function is used to describe the suppression effect of the current PID control parameters on the first vibration signal.

[0072] The loss function is obtained by dividing the value of the vibration frequency domain signal by a certain value, taking the exponent, and then summing the results. The exponential function exhibits a property of rapidly increasing on the right half of the real axis. When the vibration reaches the control requirements, the value of the loss function will basically stabilize around a certain value. However, when a small vibration occurs, the value of the loss function will rise sharply.

[0073] like Figure 3 As shown, in one embodiment, the soft robotic arm control method provided by the present invention further includes the following steps:

[0074] Step S310: Based on the first data and the first image, obtain the first state of the object being grasped, where the first state is the current state of the robotic arm.

[0075] Specifically, the host computer obtains the current state of the soft manipulator based on the pressure data measured by the pressure sensor on the soft manipulator, the motion data measured by the IMU sensor on the soft manipulator, and the image data inside the manipulator measured by the image sensor inside the soft manipulator.

[0076] The current state of the soft robotic arm includes the pressure sensor data, IMU sensor data, and the volume distribution inside the soft robotic arm at this moment.

[0077] Step S320: When the object being grasped is in the first state and the value of the corresponding loss function exceeds the first threshold, the first state is fitted by a convolutional neural network to obtain the first behavior. The first behavior is the best behavior of the robot when the loss function exceeds the first threshold in the first state.

[0078] Specifically, when the value of the loss function corresponding to the object being grasped exceeds the first threshold, it indicates that the soft robotic arm is vibrating significantly. The host computer will then use a convolutional neural network to fit the pressure sensor data, IMU sensor data, and volume distribution inside the soft robotic arm to obtain the optimal behavior of the soft robotic arm at this moment, i.e., the first behavior. By increasing the PID control parameters, the air intake and exhaust volume of the air pump are adjusted more quickly to rapidly reduce the vibration of the soft robotic arm, enabling the soft robotic arm to quickly and stably grasp the object being grasped.

[0079] Step S330: When the object being grasped is in the first state and the value of the corresponding loss function is lower than the second threshold, the first state is fitted by a convolutional neural network to obtain the second behavior. The second behavior is the best behavior of the robot when the loss function is lower than the second threshold in the first state.

[0080] Specifically, when the value of the loss function corresponding to the grasped object is lower than the second threshold, it indicates that the vibration of the soft manipulator is relatively small. The host computer will then use a convolutional neural network to fit the pressure sensor data, IMU sensor data, and volume distribution inside the soft manipulator at this moment to obtain the optimal behavior of the soft manipulator, i.e., the second behavior. By adjusting the PID control parameters, the soft manipulator slowly enters a stable grasping state, preventing damage to fragile objects caused by excessive force due to drastic changes in control parameters.

[0081] like Figure 4 As shown, in one embodiment, the soft robotic arm control method provided by the present invention uses a convolutional neural network to identify the first image after image edge detection processing to obtain volume distribution data at different locations inside the robotic arm. Prior to this, the method further includes the following steps:

[0082] Step S410: Acquire image data before and after the gripping motion of the robotic arm.

[0083] Specifically, the host computer uses image sensors to acquire image data of the soft robotic arm before and after it stably grasps the object being grasped.

[0084] Step S420: Perform edge detection on the image data before and after the gripping motion of the robotic arm to obtain the image to be identified after edge detection.

[0085] Specifically, the host computer performs edge detection on the image data of the soft robotic arm before and after the movement of the grasped object obtained in step S410, so as to obtain the images to be identified of different parts inside the soft robotic arm.

[0086] Step S430: The image to be identified is identified by a convolutional neural network to obtain volume distribution data before and after the robotic arm's grasping motion.

[0087] Specifically, the host computer uses a convolutional neural network to identify the image to be identified obtained in step S420, in order to obtain the volume distribution data of the soft manipulator before and after grasping the object. The deep reinforcement learning algorithm is then used to comprehensively analyze and process the volume distribution data inside the soft manipulator, thereby obtaining more flexible PID control parameters to maintain the stable grasp of the object by the soft manipulator.

[0088] like Figure 5 As shown, in one embodiment, the soft robotic arm control method provided by the present invention calls a deep reinforcement learning algorithm to analyze and process the second data and volume distribution data, and then includes the following steps:

[0089] Step S510: Call the PID algorithm to calculate the analysis and processing results of the deep reinforcement learning algorithm to obtain the pulse width modulation parameters.

[0090] Specifically, the host computer calls the PID algorithm to calculate the results obtained by the deep reinforcement learning algorithm through comprehensive analysis and processing of the load distribution data corresponding to the pressure sensor data, the vibration error corresponding to the IMU sensor data, and the volume distribution data corresponding to the soft manipulator, in order to obtain the corresponding pulse width modulation parameters, i.e., PID control parameters.

[0091] Among them, the PID algorithm stands for "proportional, integral, and derivative," which is a common control algorithm that can maintain the stable grip of a soft robotic hand by adjusting the PID control parameters based on the vibration of the soft robotic hand.

[0092] In step S520, the pulse width modulation parameters are sent to the lower-level computer connected to the robotic arm via the transmission control protocol to control the robotic arm to perform grasping motion.

[0093] Specifically, the host computer sends the pulse width modulation parameters obtained in step S510 to the slave computer connected to the soft robotic arm via the transmission control protocol (TCP communication), and the slave computer directly controls the soft robotic arm to perform stable and reliable grasping movements.

[0094] In a specific embodiment, the present invention provides a soft robotic arm control method, firstly, combining... Figure 6 As shown, the soft manipulator system controlled by this method mainly consists of three parts: a pneumatic end effector, a lower-level computer, and a higher-level computer. The pneumatic end effector comprises a soft manipulator, a flexible pressure sensor with an embedded mixed-position metal-carbon nanofilm, a built-in vision sensor, and an IMU accelerometer.

[0095] In this embodiment, the soft robotic hand consists of three soft fingers, a support structure, and a pneumatic drive system. The soft fingers are fixed by the support structure, and the pneumatic drive system is connected to the soft fingers via two air inlets and outlets. The soft fingers are made of different types of silicone. The outer silicone layer uses a slit-by-slit arrangement, allowing the soft fingers to bend more when inflated by the air pump. The inner silicone layer is coated with alternating black and white pigments to facilitate the visual sensor's recognition of the internal image of the soft robotic hand. One end of the support structure engages with the robotic arm to enable the movement and rotation of the soft robotic hand. The end of the soft robotic hand is equipped with a edge-embedded mixed-site metal carbon nanotube thin film pressure sensor, which can accurately monitor the contact force between the soft fingers and the grasped object during the grasping process. By analyzing the magnitude of the force and the frequency of vibration, the inflation volume of the air pump is adjusted to eliminate vibration.

[0096] It should be noted that the edge-embedded mixed-site metal-carbon nanofilm pressure sensor at the end effector of the soft robotic arm primarily utilizes magnetron sputtering and ultraviolet lithography techniques. A 70nm thick titanium alloy electrode is deposited on the left side of the stretchable mixed-site metal-carbon nanofilm sensor as the negative electrode, and a 70nm thick titanium alloy electrode is deposited on the right side as the positive electrode. The effective bending angle of the mixed-site metal-carbon nanofilm pressure sensor is 0° to 180°. Under the external force generated by the bending of the soft finger, the mixed-site metal-carbon nanofilm sensor undergoes strain, causing dislocations to bond between the metal nanoparticles and graphene in the carbon film. This reduces the conductive pathway, leading to an increase in the sensor's resistance. However, due to the good conductivity of the mixed-site metal-carbon nanoparticles and the increased conductivity from the conductive polyaniline particles, the conductive network can be maintained. The change in sensor resistance will cause a corresponding change in the voltage signal, thus reflecting the pressure distribution during the bending of the soft finger.

[0097] In this embodiment, the tip of the soft finger is equipped with an air pump input / output port and a vision sensor. The vision sensor acquires images of the inside of the soft finger, which are then analyzed by a host computer to determine the finger's volume distribution. The vision sensor consists of a three-color soft finger cavity and an industrial endoscope camera module. The industrial endoscope camera module is used to observe the cavity of the soft finger; an EZ-EN45S-R endoscope can be selected as the camera. Its function is to probe deep into tiny channels and observe the internal spatial structure and state within a sealed cavity. The lower-level computer is mainly used to control the air pump, acquire resistance array and vibration signals, and communicate with the host computer. It consists of a microcontroller, a piezoresistive scanning circuit, a filtering circuit, a microcontroller analog-to-digital converter, and a communication module. The microcontroller controls the air pump's airflow output and the piezoresistive scanning circuit scans the voltage distribution of the pressure sensor. The filtering circuit removes high-frequency noise from the piezoresistive scanning circuit, the microcontroller analog-to-digital converter converts the voltage signal into a digital signal, and the communication interface enables communication with the host computer, realizing pressure information acquisition and air pump airflow control. In addition, the microcontroller controls the on / off state of the piezoresistive scanning circuit to realize the row and column block scanning acquisition of the resistors, and saves the resistance data of each row scan in a variable. The high-frequency noise in the acquisition signal is filtered by using a seventh-order Butterworth filter, and the signal of the IMU sensor is acquired and its data is sent to the host computer along with the resistor array.

[0098] Combination Figure 7 As shown, after receiving pressure sensor data and IMU sensor data from the soft robotic arm, the lower-level computer sends them to the upper-level computer via TCP communication (Transmission Control Protocol). The upper-level computer processes the pressure sensor data to obtain the load distribution corresponding to the current pressure sensor data. The upper-level computer processes the IMU (Inertial Measurement Unit) sensor data to obtain the vibration error of the soft robotic arm. Furthermore, the upper-level computer acquires image data measured by the image sensor inside the soft robotic arm and performs image edge detection. Then, it uses a convolutional neural network to recognize the image to obtain the volume distribution inside the soft robotic arm. Subsequently, a DQN (Reinforcement Learning Neural Network) is used to comprehensively process the load distribution data, vibration error, and gradient distribution. After processing, a PID algorithm is called to automatically derive the current air pump vibration PWM (Pulse Width Modulation) parameters. After calculating the air pump output using the PID algorithm, the control signal is sent to the lower-level computer via TCP communication to adjust the air pump output.

[0099] In this embodiment, the image signals acquired from the visual sensor need to undergo image edge detection preprocessing before being transmitted to the deep neural network. The deep neural network uses Gaussian filtering to smooth the image, adaptive illumination contrast correction, image binarization, finding the image intensity gradient, applying non-maximum suppression techniques to eliminate false edge detections, and a double thresholding method to determine image boundaries, and uses hysteresis techniques to track the boundaries. After image edge detection processing of the visual sensor signals, a CNN convolutional neural network accurately identifies the volume distribution of the soft finger before it begins to move after grasping, and transmits this data to the subsequent neural network for processing during movement. To train the network, an IMU accelerometer is used to acquire the vibration signal of the grasped object and evaluate the effect of the vibration. The IMU accelerometer is placed at the point of contact between the grasped object and the soft manipulator, and its frequency domain is analyzed after anti-aliasing filtering, AD conversion, discrete fast Fourier transform, and windowing.

[0100] In this embodiment, to describe the vibration suppression effect, the value of the frequency domain signal is divided by a certain value, the exponent is taken, and then the results are summed to obtain the vibration control loss function: The exponential function exhibits a sharp increase on the right half of the real axis. When vibration reaches the control requirements, the value of the loss function stabilizes around a certain level. However, when minor vibrations occur, the value of the loss function rises sharply. The processed load distribution data and volume distribution data are transmitted to the DQN reinforcement learning neural network to automatically derive the current air pump vibration PID control parameters, adjusting for different gripping conditions. The DQN reinforcement learning neural network requires the current state S and reward R. The network then fits a Q-table to derive the optimal behavior A under the current state S. Let the current state S represent the robot's position, velocity, acceleration, and the volume distribution data of the soft fingers collected by the vision sensor before movement and the finger load distribution collected by the flexible pressure sensor. When the loss function value is greater than a set threshold (i.e., vibration exceeds a certain level), R is set to 0. When the loss function value is less than the set threshold (i.e., vibration falls below a certain level), R is set to 1. The optimal behavior A is used to adjust the three PID parameters based on the magnitude of the loss function value.

[0101] The host computer controls the air pump using a PID algorithm, where the error is the difference between the pressure distribution load of the flexible pressure sensor and its stable value. After the host computer calculates the error using the PID algorithm, the control signal is sent to the air pump via the lower-level microcontroller to control the air intake and exhaust volume, thereby controlling the vibration.

[0102] The aforementioned soft robotic arm control method involves the host computer acquiring data from a vision sensor before the robotic arm stabilizes and begins movement. This data is transmitted to the host computer to determine the current internal state of the fingers and is accurately identified using a convolutional neural network to determine the volumetric spatial distribution of the grasping object. Simultaneously, pressure sensor data is used to comprehensively identify the current state of the soft fingers. When the robotic arm system begins to move, the host computer shuts down the vision sensor. The lower-level computer collects pressure data from the flexible pressure sensor and feeds it back to the host computer in real time. Based on the transmitted data, the host computer calculates the optimal PID control parameters using a DQN reinforcement learning neural network and transmits the control signal to the lower-level computer in real time. This continuously adjusts the air intake of the air pump to synchronize with the vibration of the soft fingers, achieving adaptive control and thus enabling reliable and stable grasping of fragile objects.

[0103] The soft robotic arm control device provided by the present invention will be described below. The soft robotic arm control device described below can be referred to in correspondence with the soft robotic arm control method described above.

[0104] like Figure 8 As shown, in one embodiment, a soft robotic arm control device includes a data receiving module 810, a first processing module 820, an image recognition module 830, and a second processing module 840.

[0105] The data receiving module 810 is used to receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm.

[0106] The first processing module 820 is used to process the first data to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0107] The image recognition module 830 is used to recognize the first image after image edge detection processing through a convolutional neural network in order to obtain volume distribution data at different locations inside the robot arm.

[0108] The second processing module 840 is used to call a deep reinforcement learning algorithm to analyze and process the second data and volume distribution data to obtain the first control parameter. The first control parameter is a PID control parameter used to control the air output of the air pump connected to the robot arm.

[0109] In this embodiment, the soft robotic arm control device provided by the present invention further includes a vibration signal processing module, used for:

[0110] Based on IMU sensor data, a first vibration signal is acquired, which is the vibration signal of the object being grasped by the robotic arm.

[0111] Based on the first vibration signal, a loss function corresponding to the vibration error is obtained. The loss function is used to describe the suppression effect on the first vibration effect.

[0112] In this embodiment, the soft robotic arm control device provided by the present invention further includes an image processing module, used for:

[0113] Acquire image data before and after the robotic arm's grasping motion.

[0114] Edge detection is performed on the image data before and after the robotic arm's grasping motion to obtain the image to be identified after edge detection.

[0115] In this embodiment, the image recognition module of the soft robotic arm control device provided by the present invention is specifically used for:

[0116] A convolutional neural network is used to identify the image to obtain volume distribution data before and after the robotic arm's grasping motion.

[0117] In this embodiment, the soft robotic arm control device provided by the present invention further includes a PID algorithm processing module, used for:

[0118] The PID algorithm is invoked to calculate the pulse width modulation parameters by analyzing the results of the deep reinforcement learning algorithm.

[0119] The pulse width modulation parameters are sent to the lower-level computer connected to the robotic arm via a transmission control protocol to control the robotic arm to perform grasping movements.

[0120] In this embodiment, the soft robotic arm control device provided by the present invention further includes an optimal behavior acquisition module, used for:

[0121] Based on the first data and the first image, the first state of the object being grasped is obtained, which is the current state of the robotic arm.

[0122] When the object being grasped is in the first state and the value of the corresponding loss function exceeds the first threshold, the first state is fitted by a convolutional neural network to obtain the first behavior. The first behavior is the best behavior of the robot when the loss function exceeds the first threshold in the first state.

[0123] When the object being grasped is in the first state and the corresponding loss function value is lower than the second threshold, the first state is fitted by a convolutional neural network to obtain the second behavior. The second behavior is the best behavior of the robot when the loss function is lower than the second threshold in the first state.

[0124] The present invention also provides a soft robotic arm, controlled by any of the soft robotic arm control methods described above, comprising:

[0125] The robotic arm consists of soft fingers, a support structure, and pneumatic drive tubes. The support structure is used to support and fix multiple soft fingers, and the pneumatic drive tubes are connected to the soft fingers and control the extension and bending of the soft fingers by air intake and exhaust.

[0126] It should be noted that pressure sensors are installed on the parts of the robotic arm that come into contact with the object being grasped, to acquire pressure during the grasping motion. Additionally, an IMU sensor is installed on the robotic arm to collect motion data such as position, speed, and acceleration during the grasping motion. An image sensor or vision sensor is installed inside the robotic arm to monitor the volume changes of the soft material at different locations on the robotic arm before and after being compressed by the grasping motion. A robotic arm is connected to one end of the robotic arm, which is connected to a support structure and used to rotate and move the robotic arm to grasp the object.

[0127] The lower-level machine is electrically connected to the robot body and is used to receive sensor data from the robot body and directly control the gripping motion of the robot body by controlling the air intake and exhaust of the air drive tube.

[0128] It should be noted that the lower-level computer is electrically connected to the pressure sensor, IMU sensor, and image sensor on the robot body to obtain the corresponding sensor data of the pressure sensor, IMU sensor, and image sensor under different motion states of the robot body.

[0129] The host computer, electrically connected to the slave computer, is used to receive sensor data from the slave computer via a transmission control protocol and process the sensor data to obtain control parameters for controlling the gripping motion of the robotic arm body, and send the control parameters to the slave computer via the transmission control protocol.

[0130] Figure 9 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 9 As shown. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a soft robotic arm control method, which includes:

[0131] Receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm. The first image is the image data measured by the image sensor inside the robot arm.

[0132] The first data is processed to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0133] The first image after edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm.

[0134] A deep reinforcement learning algorithm is invoked to analyze and process the second data and volume distribution data to obtain the first control parameter, which is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

[0135] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0136] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements a soft robotic arm control method, the method comprising:

[0137] Receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm. The first image is the image data measured by the image sensor inside the robot arm.

[0138] The first data is processed to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0139] The first image after edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm.

[0140] A deep reinforcement learning algorithm is invoked to analyze and process the second data and volume distribution data to obtain the first control parameter, which is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

[0141] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and when the processor executes the computer instructions, it implements a soft robotic arm control method, the method comprising:

[0142] Receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm. The first image is the image data measured by the image sensor inside the robot arm.

[0143] The first data is processed to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data.

[0144] The first image after edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm.

[0145] A deep reinforcement learning algorithm is invoked to analyze and process the second data and volume distribution data to obtain the first control parameter, which is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0147] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A control method for a soft robotic arm, characterized in that, The method includes: Receive first data and first image, wherein the first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm; The first data is processed to obtain the second data, which is the load distribution data corresponding to the pressure sensor data and the vibration error corresponding to the IMU sensor data. The first image after edge detection is identified by a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm. The second data and volume distribution data are analyzed and processed using a deep reinforcement learning algorithm to obtain a first control parameter. The first control parameter is a PID control parameter used to control the air intake and output of the air pump connected to the robotic arm.

2. The soft robotic arm control method according to claim 1, characterized in that, The process of processing the first data to obtain the second data further includes, prior to: Based on the IMU sensor data, a first vibration signal is obtained, wherein the first vibration signal is the vibration signal of the object being grasped by the robotic arm; Based on the first vibration signal, a loss function corresponding to the vibration error is obtained, and the loss function is used to describe the suppression effect on the first vibration signal.

3. The soft robotic arm control method according to claim 2, characterized in that, The method further includes: Based on the first data and the first image, the first state of the grasped object is obtained, and the first state is the current state of the robotic arm; When the object being grasped is in a first state and the value of the corresponding loss function exceeds a first threshold, the convolutional neural network is invoked to fit the first state to obtain a first behavior. The first behavior is the best behavior of the robotic arm when the loss function exceeds the first threshold in the first state. The first action is used to increase the first control parameter.

4. The soft robotic arm control method according to claim 3, characterized in that, The method further includes: When the object being grasped is in a first state and the value of the corresponding loss function is lower than a second threshold, the convolutional neural network is invoked to fit the first state to obtain a second behavior, which is the best behavior of the robotic arm when the loss function is lower than the second threshold in the first state. The second action is used to reduce the first control parameter.

5. The soft robotic arm control method according to claim 1, characterized in that, The step of identifying the first image after edge detection processing using a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm also includes: Acquire image data before and after the robotic arm's grasping motion; Edge detection is performed on the image data before and after the grasping motion of the robotic arm to obtain the image to be identified after edge detection.

6. The soft robotic arm control method according to claim 5, characterized in that, The step of recognizing the first image after edge detection processing using a convolutional neural network to obtain volume distribution data at different locations inside the robotic arm includes: The convolutional neural network is invoked to identify the image to be recognized in order to obtain the volume distribution data of the robotic arm before and after the grasping motion.

7. The soft robotic arm control method according to claim 1, characterized in that, The process of invoking a deep reinforcement learning algorithm to analyze and process the second data and the volume distribution data includes: The PID algorithm is invoked to calculate the pulse width modulation parameters by analyzing the results of the deep reinforcement learning algorithm. The pulse width modulation parameters are sent to the lower-level computer connected to the robotic arm via a transmission control protocol to control the robotic arm to perform grasping movements.

8. A soft robotic arm control device, characterized in that, The device includes: The data receiving module is used to receive first data and first image. The first data is the pressure sensor data and IMU sensor data corresponding to the robot arm, and the first image is the image data measured by the image sensor inside the robot arm. A first processing module is used to process the first data to obtain second data, wherein the second data is load distribution data corresponding to the pressure sensor data and vibration error corresponding to the IMU sensor data; The image recognition module is used to recognize the first image after image edge detection processing through a convolutional neural network in order to obtain volume distribution data at different locations inside the robot arm; The second processing module is used to call a deep reinforcement learning algorithm to analyze and process the second data and volume distribution data to obtain a first control parameter. The first control parameter is a PID control parameter used to control the air output of the air pump connected to the robotic arm.

9. A soft robotic hand, characterized in that, Controlled by the soft robotic arm control method according to any one of claims 1 to 7, including: The robotic arm body consists of soft fingers, a support structure, and an air drive tube. The support structure is used to support and fix multiple soft fingers. The air drive tube is connected to the soft fingers and controls the extension and bending of the soft fingers by air intake and exhaust. The lower-level machine is electrically connected to the robot body and is used to receive sensor data from the robot body and directly control the gripping motion of the robot body by controlling the air intake and exhaust of the pneumatic drive tube. The host computer, electrically connected to the slave computer, is used to receive sensor data from the slave computer via a transmission control protocol and process the sensor data to obtain control parameters for controlling the gripping motion of the robotic arm body, and send the control parameters to the slave computer via the transmission control protocol.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images