Cross-mode cooperative control method for multi-cavity pneumatic soft gripper of picking robot

Abstract

The application discloses a cross-mode cooperative control method of a multi-cavity pneumatic soft hand claw of a picking robot, and comprises the following steps: generating a pre-grasping and releasing pose through visual positioning; moving to the pre-grasping pose and establishing a cavity direction mapping from local side swing to global circumferential direction; initializing and updating a coupling model on line in a grasping mode, performing a constrained predictive grasping; in a twisting mode, activating the same-direction side swing cavity with a phase shift, and jointly adjusting the inward bending cavity compensation to compensate for the normal unloading; unifying the direction, normalizing the scale and continuously accumulating four types of residual errors, fusing single-channel and channel cooperative evidence to determine the fruit stem separation; after the separation, the fruit is carried to the release pose, and the pressure is sequentially released after the tolerance is met. The application can solve the problems that the existing technology is disconnected from the grasping control and visual positioning, the side swing deformation disturbance normal contact force causes unloading or overpressure, and the fruit stem is prematurely triggered, so that the fruit stem twisting separation is effectively realized.

Description

Technical Field

[0001] This invention relates to the field of electronic information technology, and more specifically to a cross-mode collaborative control method for a multi-cavity pneumatic soft gripper of a harvesting robot. Background Technology

[0002] Pneumatic soft grippers utilize the flexibility of elastic materials to envelop and grasp fragile fruits. However, in practical applications, the nonlinearity and hysteresis effects of the elastic materials, pressure coupling between multiple cavities, and variations in the size of the grasped fruit and the contact position can all cause deviations in the contact force under fixed air pressure conditions. Using simple open-loop air pressure control or fixed-parameter force control alone makes it difficult to simultaneously satisfy finger-by-finger grasping force tracking, three-finger force balance, and pressure safety constraints.

[0003] For fruits that require twisting to separate the stem, using a robotic arm that rotates the entire arm or adds a rigid shearing mechanism to perform the shearing will increase the difficulty of motion planning, end-effector quality, and the risk of accidental injury. When using the lateral swing cavity of the soft fingers to generate torque, the lateral swing deformation will disturb the normal contact force, causing partial unloading of the fingers, fruit slippage, or localized overpressure.

[0004] Current methods for determining fruit stem separation often rely on sudden drops in contact force or single posture thresholds. However, transient force or displacement signals generated by short-term slippage, robotic arm vibration, and localized contact are similar to the characteristics of actual fruit stem separation, which can easily lead to premature misjudgment of fruit stem separation and trigger release.

[0005] Furthermore, if the target position is given at the beginning of the operation but the visual result is not converted into the robot's pre-grasping pose and fruit basket release pose, the perception information will be disconnected from the pneumatic gripping control. If the pressure is released immediately after the fruit stem is separated without holding the fruit and moving it above the basket opening, there is a risk that the fruit will fall into a non-target position. Summary of the Invention

[0006] In view of this, the present invention provides a cross-mode collaborative control method for a multi-cavity pneumatic soft gripper of a harvesting robot, in order to solve the problems of disconnect between grasping control and visual positioning, unloading or overpressure caused by lateral deformation disturbance of normal contact force, and premature false triggering in the existing technology, thereby effectively realizing the torsion separation of fruit stems.

[0007] A cross-mode cooperative control method for a multi-cavity pneumatic soft gripper of a harvesting robot, wherein the multi-cavity pneumatic soft gripper includes a plurality of soft fingers spaced apart circumferentially along the gripper base, each soft finger having at least an independent inwardly bending cavity group, a left-side swing cavity group, and a right-side swing cavity group, each cavity group being connected to a pressure regulating channel, the method comprising the following steps: S1: Use depth vision to obtain spatial geometric information of the fruit and basket opening, transform it to the robot coordinate system according to the hand-eye calibration relationship, and generate pre-grabbing pose and release pose; S2, control the robot to move to the pre-grasping pose and start sensor data acquisition, and establish a cavity direction mapping between the local lateral swing direction and the global circumferential direction of the gripper based on the installation orientation of the soft finger on the gripper base; S3, in grasping mode, initialize the lateral unloading coefficient of each soft finger with offline calibration value, update the inward force increase coefficient and contact force increment bias online under the condition that the lateral pressure increment is zero, and use the coupled prediction model containing the normal force increase term formed by the inward bending pressure increment and the normal unloading term caused by the lateral pressure increment to predict the finger-by-finger contact force, select and execute the pressure increment in predictive control mode until the finger-by-finger contact force meets the stability condition. S4. In the twisting and picking mode, the same-direction side-swing cavity group that generates the same global circumferential torque on the target fruit is selected according to the cavity direction mapping. They are activated sequentially in a phase-staggered time sequence. When the side-swing pressure excitation is effective, the side-swing unloading coefficient is updated. The current phase side-swing pressure increment is combined with the compensation pressure increment of the three inward bending cavity groups to form a candidate control quantity, so as to compensate for the finger-by-finger normal unloading caused by the side-swing while generating continuous circumferential torque. S5, during the execution of S4, unifies the direction, normalizes the scale, and continuously accumulates the contact force drop residual, base reaction angular velocity residual, contact force prediction residual, and slip residual, and integrates single-channel evidence and channel collaborative evidence to determine the separation of fruit stem; S6, in response to the fruit stem separation determination, to maintain pressure and hold the target fruit between several soft fingers, control the robot to move to the release pose, and after meeting the release tolerance, switch to the release mode and release pressure sequentially.

[0008] The cross-mode cooperative control method for a multi-cavity pneumatic soft gripper of a harvesting robot provided by the present invention has the following beneficial effects: 1. This invention utilizes depth vision to simultaneously generate pre-grasping pose and fruit basket release pose, switching between in-position tolerance constraint modes to transfer the fruit after stem separation to the release pose before depressurization. This allows the visual results to directly participate in the entire process of robot approach, grasping, transporting, and releasing the fruit, avoiding the disconnect between grasping control and visual positioning caused by the camera only providing isolated position data. This invention achieves closed-loop collaborative control of the entire process from positioning, grasping, twisting, transporting to fixed-point release.

[0009] 2. This invention establishes a local-global cavity direction mapping. Based on this mapping, local left-side or right-side swing cavity groups are selected from three circumferentially mounted soft fingers to generate a consistent global circumferential torque. This avoids the torque directions canceling out or conflicting due to simply applying pressure to local cavities with the same name. At the same time, the fruit stalk can be twisted and separated without rotating the entire robotic arm.

[0010] 3. This invention establishes a coupled prediction model that includes an inward force amplification term and a lateral unloading term, and updates the model parameters online. During twisting and unloading, the upper limit of the lateral pressure is constrained by the real-time gripping margin and compensation capability. The uncompensated unloading error term in the candidate evaluation function guides the controller to prioritize the control quantity that can be compensated by the inward force amplification for the lateral unloading amount. Thus, within the same model, cost, and constraint system, the circumferential drive and normal holding are coordinated, compensating for the finger-by-finger normal unloading caused by the lateral swing, and avoiding unloading or overpressure caused by the normal contact force of the lateral swing deformation disturbance.

[0011] 4. This invention first unifies the direction and scale of the four types of residuals according to their respective operating baselines, then performs bounded continuous accumulation on each, and strengthens the multi-source consistency evidence for the synergistic terms through channels; the judgment is only output after the separation risk persistence condition and the threshold conditions of at least two channels are met, so that a sudden drop in contact force, short-term slippage or local vibration disturbance will not trigger the separation judgment alone, and the instantaneous disturbance is distinguished from the true continuous separation from the mechanism, which solves the problem of premature false triggering caused by the similarity between short-term slippage, vibration and the true fruit stem separation characteristics.

[0012] 5. To address the problem of fruit falling to non-target positions due to immediate pressure release after fruit stem separation, this invention sets up a holding and transporting stage between the separation determination and the release action. This stage maintains pressure to keep the target fruit in place and is then transported to the release position by a robot. Pressure is only released sequentially after the relative positions of the basket openings meet the release tolerance, thereby closing the complete control chain from fruit stem separation to fixed-point fruit release and preventing fruit from falling, being damaged, or scattering. Attached Figure Description

[0013] Figure 1 A schematic diagram of the structure of a multi-cavity pneumatic soft gripper; Figure 2 A schematic cross-sectional view of the multi-cavity structure of a single soft finger; Figure 3 A flowchart illustrating the cross-mode collaborative control method for a multi-cavity pneumatic soft gripper of a harvesting robot; Figure 4 Equivalent simulation curves for gripping force and three-finger force difference; Figure 5 Equivalent simulation curves for phase-shifted side swing and normal compensation in torsion-disconnect mode; Figure 6 The equivalent simulation curves for the four-channel residuals and the risk of stem separation are shown. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0015] Embodiments of the present invention provide a cross-mode cooperative control method for a multi-cavity pneumatic soft gripper of a harvesting robot. Please refer to [link to relevant documentation]. Figure 1 and Figure 2 The multi-cavity pneumatic soft gripper includes three soft fingers 20 spaced apart along the circumference of the gripper base. The soft fingers 20 are mounted on the gripper base 10 installed at the end of the robotic arm.

[0016] Each soft finger 20 has at least three independent inwardly bending cavity groups 21, left-side swing cavity groups 23, and right-side swing cavity groups 24, each cavity group being connected to a pressure regulating channel. Each soft finger also has an outwardly bending cavity group 22.

[0017] The soft finger 20 has a contact force sensor 31 and a slip sensor 32 on its inner side, which respectively collect the normal contact force and the relative slip amount. The depth camera 50 is a depth camera installed at the end of the robot, and its field of view covers the target fruit and the opening of the fruit basket.

[0018] The soft fingers 20 are integrally molded from silicone material. Each soft finger 20 contains ten isolated pneumatic chambers. Three of the pneumatic chambers are interconnected to form the inward bending chamber group 21, three of the pneumatic chambers are interconnected to form the outward bending chamber group 22, two of the pneumatic chambers are interconnected to form the left swing chamber group 23, and the other two pneumatic chambers are interconnected to form the right swing chamber group 24. The three soft fingers 20 are arranged at 120° intervals to form twelve chamber groups that can independently intake, maintain pressure, and exhaust air.

[0019] Each cavity group has an air pressure sensor in its inlet pipe, and an inertial measurement unit is installed on the gripper base 10. The harvesting robot also includes a controller and a pressure regulation module that controls the air intake, pressure holding and exhaust of the twelve cavity groups.

[0020] Please see Figure 3 The method includes steps S1 to S6: S1 uses depth vision to acquire spatial geometric information of the fruit and basket opening, transforms it to the robot coordinate system according to the hand-eye calibration relationship, and generates pre-grabbing pose and release pose.

[0021] Specifically, step S1 includes: The grasping center and approach direction of the target fruit, as well as the center position and normal direction of the fruit basket opening, are obtained through the visual positioning component. Based on the hand-eye calibration relationship, the grasping center, approach direction, center position of the basket opening, and normal direction are transformed into the robot coordinate system to generate the pre-grab pose of the target fruit and the release pose above the fruit basket. The visual positioning component includes a depth camera and a visual processing module. The visual processing module determines the grasping center, fruit size, and approach direction based on the color image and depth data of the target fruit, and determines the center position of the basket opening, the normal direction of the basket opening, and the range of the basket opening based on the edge point set of the fruit basket. After the position error and attitude error between the robot end effector and the pre-grasping pose or release pose enter the preset tolerance range, the mode supervision unit is allowed to enter the corresponding grasping mode or release mode.

[0022] S2, control the robot to move to the pre-grasping pose and start sensor data acquisition, and establish the cavity direction mapping between the local lateral swing direction and the global circumferential direction of the gripper based on the installation orientation of the soft finger on the gripper base.

[0023] In this step, the robot is controlled to move the multi-cavity pneumatic soft gripper to the pre-grasping pose and initiate sensor data acquisition. In the gripping mode, twisting mode, and holding and transporting mode, the contact force of the three soft fingers, the cavity pressure of each cavity group, the twisting reaction angular velocity characteristics of the gripper base around the axis, and the slippage of each soft finger are continuously collected. Based on the installation orientation of the three soft fingers on the gripper base, a cavity orientation mapping between the local lateral swing direction and the global circumferential direction of the gripper is established, and the active cavity group, holding cavity group, exhaust cavity group, and pressure and contact force constraints are determined for the gripping mode, twisting mode, holding and transporting mode, and release mode.

[0024] The cavity orientation mapping records the installation angle of each soft finger, the global circumferential symbol corresponding to the left swing cavity group, and the global circumferential symbol corresponding to the right swing cavity group. In the twisting and untwisting mode, only the side swing cavity groups whose global circumferential symbols are consistent with the target twisting and untwisting direction are selected, and the unselected side swing cavity groups are kept in the exhaust state or low pressure holding state.

[0025] S3, in grasping mode, initialize the lateral unloading coefficient of each soft finger with offline calibration value, update the inward force increase coefficient and contact force increment bias online under the condition that the lateral pressure increment is zero, and use the coupled prediction model containing the normal force increase term formed by the inward bending pressure increment and the normal unloading term caused by the lateral pressure increment to predict the finger-by-finger contact force, select and execute the pressure increment in predictive control mode until the finger-by-finger contact force meets the stability condition.

[0026] In this step, under the grasping mode, three inwardly bending cavity groups are used as active cavity groups. The lateral unloading coefficient of each soft finger is initialized with offline calibration values, and the inward force amplification coefficient and contact force increment bias are updated online under the condition that the lateral pressure increment is zero. The coupled prediction model represents the normal force amplification caused by the inward bending pressure increment and the normal unloading caused by the same-direction lateral pressure increment, respectively. Candidate control quantities are formed from a finite set of pressure increments, and the finger-by-finger contact force corresponding to each candidate control quantity is predicted. The finger-by-finger target force tracking error, the three-finger contact force imbalance, and the control increment constitute the grasping cost. Under the constraints of finger-by-finger safe grasping force, cavity pressure, and cavity pressure change rate, the candidate control quantity with the minimum cost is selected and its first pressure increment is executed until the finger-by-finger contact force meets the stability condition.

[0027] The coupled prediction model performs a one-step prediction according to the following formula:

[0028] in, For the first The soft finger is from the first Control cycle prediction Controlling the contact force during the cycle, For the first Control cycle number Measured contact force of a soft finger For the first The inward force amplification coefficient of the control cycle, For the first The pressure increment of the inward bending cavity assembly during the control cycle. For the first Side-slip unloading coefficient during control period For the first The pressure increment of the side-swing chamber assembly during the control cycle. For the first Control cycle contact force increment bias; The lateral swing unloading coefficient is initialized using offline calibration values; in gripping mode, the lateral swing pressure increment is zero, the lateral swing unloading coefficient remains at its initial value, and the inward force increment coefficient and contact force increment bias are updated online according to the following formula; and after entering step S4 and generating an effective lateral swing excitation, the lateral swing unloading coefficient is also corrected according to the following formula:

[0029] in, For the first The first soft finger The parameter vector of the control cycle includes, in turn, the inward boosting coefficient. Side-slip unloading coefficient and contact force increment bias ; For the first The first soft finger The parameter vector for controlling the cycle; For the eigenvectors, sequentially from the first... Incremental pressure in the control cycle The increase in lateral pressure after taking the negative value And composed of the constant 1; The difference between the measured contact forces in two adjacent control cycles; It is a projection operator used to limit the inward force amplification coefficient, the lateral unloading coefficient, and the contact force increment bias within a preset physical feasible region. Update the gain matrix as a diagonal positive definite matrix; This is a normalization constant, greater than zero; Represents the square of the L2 norm; This indicates transpose; in gripping mode, the side-swing pressure characteristic component is zero, and the side-swing unloading coefficient remains at the initial value of offline calibration; after entering S4, the side-swing unloading coefficient is only corrected when both the side-swing pressure excitation and the contact state are effective.

[0030] S4. In the twisting and picking mode, the same-direction side-swing cavity group that generates the same global circumferential torque on the target fruit is selected according to the cavity direction mapping. They are activated sequentially in a phase-staggered time sequence. When the side-swing pressure excitation is effective, the side-swing unloading coefficient is updated. The current phase side-swing pressure increment is combined with the compensation pressure increment of the three inward bending cavity groups to form a candidate control quantity, so as to compensate for the finger-by-finger normal unloading caused by the side-swing while generating continuous circumferential torque.

[0031] In this step, the contact force state, cavity pressure state, inward force amplification coefficient, initial value of lateral swing unloading coefficient, and contact force increment bias from the end of step S3 are inherited in twist-and-pick mode. According to the cavity direction mapping, the same-direction lateral swing cavity group that generates the same global circumferential torque on the target fruit is selected from the left and right lateral swing cavity groups of each soft finger, and the three same-direction lateral swing cavity groups are activated sequentially according to the phase-staggered timing. The lateral swing unloading coefficient is updated when the lateral swing pressure excitation is effective and the contact state is effective. In each phase, the pressure increment of the currently activated same-direction lateral swing cavity group is combined with the compensation pressure increment of the three inward bending cavity groups to form a candidate control quantity. The candidate control quantity is selected in a rolling manner according to the circumferential torque proxy quantity, the three-finger normal contact force balance quantity, and the uncompensated lateral swing unloading quantity, so as to compensate for the finger-by-finger normal unloading caused by the lateral swing while generating continuous circumferential torque.

[0032] In this embodiment, the pressure increments of the three inwardly curved cavity groups in step S3 are respectively taken from a finite set. And form candidate control variables, This represents the preset discrete pressure increment basic step size; in step S4, the lateral pressure increment of the current phase and the three inward bending compensation pressure increments form candidate control quantities; the control time domain of the predictive control is 1, the first step of each candidate pressure sequence takes the candidate control quantity, the pressure increment of subsequent prediction steps is zero, and the prediction is recursively based on the cavity pressure and contact force state after the first step; The cost of the candidate stress sequence under each operating mode is calculated using the following formula:

[0033] Each error term is calculated according to the following formula:

[0034]

[0035]

[0036]

[0037] In the formula, For the first The total cost of candidate stress sequences under various work modes. For prediction in the time domain, the value range is 10 to 25 control cycles; For the first The finger-by-finger target force tracking error of the prediction step is constituted by the difference between the finger-by-finger predicted contact force and the finger-by-finger target contact force; For the first The imbalance of the three-finger contact force in the prediction step is constituted by the difference between the predicted contact force of each finger and the average predicted contact force of the three fingers. For the first The tracking error of the circumferential torque surrogate in the prediction step is constituted by the difference between the predicted circumferential torque surrogate and the target circumferential torque surrogate. For the first The side sway unloading error of the predicted step that was not compensated for by inward pressurization; For the first The soft finger in the first The predicted contact force of the prediction step is retained using the positive part operator to retain the portion of the side swing unloading amount that exceeds the inward force compensation amount; For the first Reference contact force of a soft finger; For the three soft fingers in the first The average value of the predicted contact force for the predicted step; For the first The predicted circumferential torque proxy is obtained by summing the product of the candidate pressure increment of each unidirectional side swing cavity and the offline calibrated circumferential contribution coefficient. The target circumferential torque proxy is preset based on the fruit type and fruit stalk characteristics; This is the mode switching coefficient, which is 0 in grasping mode and 1 in twisting mode. , , , , All are non-negative weights; To control the increment, To control the incremental L2 squared value, which is used to suppress pressure abrupt changes between adjacent control cycles; For the first The soft finger in the first Pressure increment of the side-swing chamber assembly during the control cycle; For the first The soft finger in the first Control the inward bending cavity pressure increment during the control cycle; The positive part operator is used to set negative values ​​to zero; each candidate pressure sequence satisfies the upper limit of cavity pressure, the upper limit of cavity pressure change rate, and the upper limit of finger contact force. The controller only executes the first pressure increment of the candidate sequence with the minimum total cost and reconstructs the candidate sequence in the next control cycle.

[0038] Each predicted quantity is based on the first... The first control cycle starting from Predictive step calculations enable the grasping mode to evaluate only the grasping quality, while the twisting and unwinding mode evaluates both the circumferential drive and unloading compensation in addition to the grasping quality.

[0039] In step S4, a side-swing cycle is formed according to the phase-staggered timing sequence. This side-swing cycle is divided into three phase time periods, and each phase time period mainly activates one unidirectional side-swing cavity group; The soft finger in the first The upper limit of the candidate pressure of the side-swing chamber during the control cycle is determined by the following formula:

[0040] in, For the first The soft finger in the first Upper limit of candidate pressure for the side-swing chamber during the control cycle. The rated pressure of the side-swing chamber assembly. For the first The soft finger in the first Controlling the side-swing chamber pressure during the control cycle For the first The minimum holding force of a soft finger, This is the maximum allowable inward compensation pressure increment for this control cycle. This is a regularization term.

[0041] Current holding capacity With the ability to add internal compensation The allowable increase in lateral pressure is calculated by converting the real-time lateral unloading coefficient. Therefore, the amplitude limit is directly constrained by the finger-by-finger gripping state, rather than using a fixed pressure ratio.

[0042] S5, during the execution of S4, unifies the direction, normalizes the scale, and continuously accumulates the contact force drop residual, base reaction angular velocity residual, contact force prediction residual, and slip residual, and integrates single-channel evidence and channel collaborative evidence to determine the separation of the fruit stem.

[0043] In this step, four types of instantaneous residual sources are defined: contact force drop channel, contact force prediction channel, slip channel, and base reaction angular velocity channel. Specifically, the contact force drop channel takes the maximum value of the contact force decrease in three adjacent control cycles of the soft fingers; the contact force prediction channel takes the maximum value of the difference between the predicted contact force and the measured contact force of three soft fingers in one step; the slip channel takes the maximum value of the slip increment in three adjacent control cycles of the soft fingers; and the base reaction angular velocity channel takes the orthaxial reaction angular velocity characteristic output by the inertial measurement unit.

[0044] The contact force prediction residual is obtained using a coupled prediction model. The contact force drop channel, contact force prediction channel, and sliding channel take the maximum value of the change in the corresponding direction of the three soft fingers, respectively. The base reaction angular velocity channel takes the angular velocity characteristic of the axial reaction output by the inertial measurement unit. The four types of instantaneous residuals are unified in direction and normalized in scale. The non-negative cumulative quantity with attenuation is updated, and the synergistic terms of each non-negative cumulative quantity and the cumulative quantities of different channels are merged into the separation risk quantity. When the separation risk quantity reaches the separation threshold for a preset number of control cycles and the non-negative cumulative quantity of at least two channels reaches the corresponding channel threshold, the fruit stem separation judgment is output and the process proceeds to step S6. Otherwise, the process returns to step S4.

[0045] In practice, the instantaneous residuals are first unified into dimensionless non-negative residuals according to the following formula:

[0046] In the formula, For channel index, , , , , These correspond to the contact force drop channel, the base reaction angular velocity channel, the contact force prediction channel, and the sliding channel, respectively. For the first The channel is in The dimensionless nonnegative residual of the control period; This is a direction coefficient used to make the residuals that are aligned with the direction of separation of the fruit stalk positive; For the first The channel is in The instantaneous residual value of the control cycle; For the first The average operating baseline of the channel; For the first The operating baseline variance of the channel; The regularization term is used; the mean and variance of the operating baseline are determined within the non-separation steady-state window after entering the twist-and-untie mode, and frozen within a single separation decision window.

[0047] Thus, four types of channels with different dimensions, different numbers of fingers, and different noise levels can be integrated into the same accumulation and fusion framework.

[0048] Then update the bounded continuous cumulative amount according to the following formula:

[0049] In the formula, For the first The channel is in The bounded cumulative amount of the control period, with a value ranging from 0 to 1; For the first The channel is in Bounded cumulative amount over the control period; This is the memory coefficient, with a value ranging from 0 to 1, used to control the decay rate of historical information; This is a non-negative drift term used to suppress slow accumulation; the accumulated amount decreases when the transient anomaly does not persist, and only the continuously existing separation direction residuals are preserved, thereby reducing premature triggering caused by short-term slippage, vibration, or local contact disturbances; then the separation risk amount is obtained by fusing according to the following formula:

[0050] In the formula, For the first Separation risk quantity during the control cycle; It is a collection of channels, including the contact force drop channel, the base reaction angular velocity channel, the contact force prediction channel, and the sliding channel; For the first Single-channel non-negative weights of the channel; For the first Channel and the Channel pairs have cooperative non-negative weights. For synergistic gain; For the first The channel is in Bounded cumulative amount over the control period; The square root term increases significantly only when at least two channels accumulate simultaneously. The controller only adjusts when the separation risk quantity is continuous. Each control cycle is not less than the separation threshold. When the bounded cumulative amount of at least two channels reaches the threshold of the corresponding channel, the fruit stem separation judgment is output. A value of 2 to 5 can be selected. If the above dual conditions are not met, return to step S4 to continue the phase misalignment process.

[0051] S6, in response to the fruit stem separation determination, to maintain pressure and hold the target fruit between several soft fingers, control the robot to move to the release pose, and after meeting the release tolerance, switch to the release mode and release pressure sequentially.

[0052] In this step, in response to the stem separation determination, the finger-by-finger target grasping force determined in step S3 is used as the holding reference, and the holding pressure is determined based on the real-time finger-by-finger contact force, inward bending cavity pressure, and finger-by-finger safe grasping force at the time of stem separation determination, so that the contact force of each soft finger is kept within the target holding range during the robot's movement; the release tolerance is determined by the center position of the basket opening, the normal direction of the basket opening, the outer size of the target fruit, and the preset safety gap, and after entering the release tolerance, the target fruit is released in the order of first stopping the side swing and then reducing the inward bending holding pressure, so that the target fruit falls into the fruit basket.

[0053] To verify the feasibility of the above method, an equivalent numerical model incorporating three-finger independent air pressure dynamics, nonlinear contact force mapping, contact condition variations, and measurement noise was established for simulation testing. Under the same conditions, the results of this invention and the traditional fixed-parameter PID method were compared. The fixed-parameter PID method and the method of this invention used the same object parameters, initial state, target value, and noise sequence. It should be noted that... Figures 4 to 6 Used to verify the trend of the corresponding control logic, but not to prove the physical calibration accuracy of each online coefficient individually.

[0054] Please see Figure 4 The control cycle of the grasping simulation is 0.02s, the prediction time domain is 18 steps, the control time domain is 1 step, the target grasping force per finger is 2.625N, the safe grasping force per finger is 3.5N, the upper limit of cavity pressure is 120kPa, and the upper limit of cavity pressure change rate is 25kPa / s. In paired simulations with 96 sets of model coefficients and random perturbations of contact conditions, the median root mean square error of the steady-state grasping force of the method of the present invention is 0.0491N, and the fixed parameter PID is 0.0664N; the median of the steady-state three-finger force imbalance of the method of the present invention is 0.0235N, and the fixed parameter PID is 0.0792N; neither method showed any exceedance of the upper limit of force or cavity pressure.

[0055] In the twisting mode, the three side-swing cavity groups are pressurized in phase, and the normal contact force is compensated by the three inward bending cavity groups. Figure 5The variation trends of phase-shifting side yaw pressure, three-finger contact force, and circumferential torque surcharge under combined compensation conditions are shown. Figure 5 It can be seen that the normal contact force can be maintained within the target range during the lateral sway drive process. It should be noted that... Figure 5 The circumferential torque in the figure is a proxy quantity used for comparison control logic and does not represent the actual torsional torque of the fruit stem.

[0056] In the random simulation of separation determination for 400 combinations of short-term slip, the premature triggering rate of the single contact force sudden drop threshold was 42.25%; within the range of simulation parameters, no premature triggering occurred in the multimodal cumulative determination, and the average effective detection delay was 0.1145s. Figure 6 Simultaneously, it displays the contact force drop, base reaction angular velocity, contact force prediction, and slippage four-channel residuals and their fusion risk quantities, from... Figure 6 It can be seen that short-term slip is not directly triggered by a single channel for release.

[0057] In summary, the cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to the above embodiments has the following beneficial effects: 1. This invention utilizes depth vision to simultaneously generate pre-grasping pose and fruit basket release pose, switching between in-position tolerance constraint modes to transfer the fruit after stem separation to the release pose before depressurization. This allows the visual results to directly participate in the entire process of robot approach, grasping, transporting, and releasing the fruit, avoiding the disconnect between grasping control and visual positioning caused by the camera only providing isolated position data. This invention achieves closed-loop collaborative control of the entire process from positioning, grasping, twisting, transporting to fixed-point release.

[0058] 2. This invention establishes a local-global cavity direction mapping. Based on this mapping, local left-side or right-side swing cavity groups are selected from three circumferentially mounted soft fingers to generate a consistent global circumferential torque. This avoids the torque directions canceling out or conflicting due to simply applying pressure to local cavities with the same name. At the same time, the fruit stalk can be twisted and separated without rotating the entire robotic arm.

[0059] 3. This invention establishes a coupled prediction model that includes an inward force amplification term and a lateral unloading term, and updates the model parameters online. During twisting and unloading, the upper limit of the lateral pressure is constrained by the real-time gripping margin and compensation capability. The uncompensated unloading error term in the candidate evaluation function guides the controller to prioritize the control quantity that can be compensated by the inward force amplification for the lateral unloading amount. Thus, within the same model, cost, and constraint system, the circumferential drive and normal holding are coordinated, compensating for the finger-by-finger normal unloading caused by the lateral swing, and avoiding unloading or overpressure caused by the normal contact force of the lateral swing deformation disturbance.

[0060] 4. This invention first unifies the direction and scale of the four types of residuals according to their respective operating baselines, then performs bounded continuous accumulation on each, and strengthens the multi-source consistency evidence for the synergistic terms through channels; the judgment is only output after the separation risk persistence condition and the threshold conditions of at least two channels are met, so that a sudden drop in contact force, short-term slippage or local vibration disturbance will not trigger the separation judgment alone, and the instantaneous disturbance is distinguished from the true continuous separation from the mechanism, which solves the problem of premature false triggering caused by the similarity between short-term slippage, vibration and the true fruit stem separation characteristics.

[0061] 5. To address the problem of fruit falling to non-target positions due to immediate pressure release after fruit stem separation, this invention sets up a holding and transporting stage between the separation determination and the release action. This stage maintains pressure to keep the target fruit in place and is then transported to the release position by a robot. Pressure is only released sequentially after the relative positions of the basket openings meet the release tolerance, thereby closing the complete control chain from fruit stem separation to fixed-point fruit release and preventing fruit from falling, being damaged, or scattering.

[0062] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. 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 all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A cross-mode cooperative control method for a multi-cavity pneumatic soft gripper of a harvesting robot, characterized in that, The multi-cavity pneumatic soft gripper includes several soft fingers spaced apart circumferentially along the gripper base. Each soft finger has at least three independent inwardly bending cavity groups, a left-side swing cavity group, and a right-side swing cavity group. Each cavity group is connected to a pressure regulating channel. The method includes the following steps: S1: Use depth vision to obtain spatial geometric information of the fruit and basket opening, transform it to the robot coordinate system according to the hand-eye calibration relationship, and generate pre-grabbing pose and release pose; S2, control the robot to move to the pre-grasping pose and start sensor data acquisition, and establish a cavity direction mapping between the local lateral swing direction and the global circumferential direction of the gripper based on the installation orientation of the soft finger on the gripper base; S3, in grasping mode, initialize the lateral unloading coefficient of each soft finger with offline calibration value, update the inward force increase coefficient and contact force increment bias online under the condition that the lateral pressure increment is zero, and use the coupled prediction model containing the normal force increase term formed by the inward bending pressure increment and the normal unloading term caused by the lateral pressure increment to predict the finger-by-finger contact force, select and execute the pressure increment in predictive control mode until the finger-by-finger contact force meets the stability condition. S4. In the twisting and picking mode, the same-direction side-swing cavity group that generates the same global circumferential torque on the target fruit is selected according to the cavity direction mapping. They are activated sequentially in a phase-staggered time sequence. When the side-swing pressure excitation is effective, the side-swing unloading coefficient is updated. The current phase side-swing pressure increment is combined with the compensation pressure increment of the three inward bending cavity groups to form a candidate control quantity, so as to compensate for the finger-by-finger normal unloading caused by the side-swing while generating continuous circumferential torque. S5, during the execution of S4, unifies the direction, normalizes the scale, and continuously accumulates the contact force drop residual, base reaction angular velocity residual, contact force prediction residual, and slip residual, and integrates single-channel evidence and channel collaborative evidence to determine the separation of fruit stem; S6, in response to the fruit stem separation determination, to maintain pressure and hold the target fruit between several soft fingers, control the robot to move to the release pose, and after meeting the release tolerance, switch to the release mode and release pressure sequentially.

2. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 1, characterized in that, Each soft finger also has an outwardly bending cavity group; each soft finger has ten mutually isolated pneumatic cavities, of which three pneumatic cavities are interconnected to form the inwardly bending cavity group, three pneumatic cavities are interconnected to form the outwardly bending cavity group, two pneumatic cavities are interconnected to form the left swing cavity group, and the other two pneumatic cavities are interconnected to form the right swing cavity group; there are three soft fingers, and the three soft fingers are arranged at 120° intervals to form twelve cavity groups that can independently intake, maintain pressure and exhaust air.

3. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 2, characterized in that, Step S1 specifically includes: The grasping center and approach direction of the target fruit, as well as the center position and normal direction of the fruit basket opening, are obtained through the visual positioning component. Based on the hand-eye calibration relationship, the grasping center, approach direction, center position of the basket opening, and normal direction are transformed into the robot coordinate system to generate the pre-grab pose of the target fruit and the release pose above the fruit basket. The visual positioning component includes a depth camera and a visual processing module. The visual processing module determines the grasping center, fruit size, and approach direction based on the color image and depth data of the target fruit, and determines the center position of the basket opening, the normal direction of the basket opening, and the range of the basket opening based on the edge point set of the fruit basket. After the position error and attitude error between the robot end effector and the pre-grasping pose or release pose enter the preset tolerance range, the mode supervision unit is allowed to enter the corresponding grasping mode or release mode.

4. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 3, characterized in that, In step S2, the cavity orientation mapping records the installation angle of each soft finger, the global circumferential symbol corresponding to the left swing cavity group, and the global circumferential symbol corresponding to the right swing cavity group. In the twist-and-untie mode, only the side swing cavity groups whose global circumferential symbols are consistent with the target twist-and-untie direction are selected, and the unselected side swing cavity groups are kept in the exhaust state or low-pressure holding state.

5. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 4, characterized in that, The coupled prediction model in step S3 performs a one-step prediction according to the following formula: in, For the first The soft finger is from the first Control cycle prediction Controlling the contact force during the cycle, For the first Control cycle number Measured contact force of a soft finger For the first The inward force amplification coefficient of the control cycle, For the first The pressure increment of the inward bending cavity assembly during the control cycle. For the first Side-slip unloading coefficient during control period For the first The pressure increment of the side-swing chamber assembly during the control cycle. For the first Control cycle contact force increment bias; The lateral swing unloading coefficient is initialized using offline calibration values; in gripping mode, the lateral swing pressure increment is zero, the lateral swing unloading coefficient remains at its initial value, and the inward force increment coefficient and contact force increment bias are updated online according to the following formula; and after entering step S4 and generating an effective lateral swing excitation, the lateral swing unloading coefficient is also corrected according to the following formula: in, For the first The first soft finger The parameter vector of the control cycle includes, in turn, the inward boosting coefficient. Side-slip unloading coefficient and contact force increment bias ; For the first The first soft finger The parameter vector for controlling the cycle; For the eigenvectors, sequentially from the first... Incremental pressure in the control cycle The increase in lateral pressure after taking the negative value And composed of the constant 1; The difference between the measured contact forces in two adjacent control cycles; It is a projection operator used to limit the inward force amplification coefficient, the lateral unloading coefficient, and the contact force increment bias within a preset physical feasible region. Update the gain matrix as a diagonal positive definite matrix; This is a normalization constant, greater than zero; Represents the square of the L2 norm; This indicates transpose; in gripping mode, the side-swing pressure characteristic component is zero, and the side-swing unloading coefficient remains at the initial value of offline calibration; after entering S4, the side-swing unloading coefficient is only corrected when both the side-swing pressure excitation and the contact state are effective.

6. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 5, characterized in that, In step S3, the pressure increments of the three inwardly curved cavity groups are respectively taken from finite sets. And form candidate control variables, This represents the preset discrete pressure increment basic step size; in step S4, the lateral pressure increment of the current phase and the three inward bending compensation pressure increments form candidate control quantities; the control time domain of the predictive control is 1, the first step of each candidate pressure sequence takes the candidate control quantity, the pressure increment of subsequent prediction steps is zero, and the prediction is recursively based on the cavity pressure and contact force state after the first step; The cost of the candidate stress sequence under each operating mode is calculated using the following formula: Each error term is calculated according to the following formula: In the formula, For the first The total cost of candidate stress sequences under various work modes. For prediction in the time domain; For the first Predictive step-by-step target force tracking error; For the first The unbalanced contact force of the three fingers in the prediction step; For the first Tracking error of circumferential torque surrogate quantity in the prediction step; For the first The side sway unloading error of the predicted step that was not compensated for by inward pressurization; For the first The soft finger in the first Predicted contact force for the predicted step; For the first Reference contact force of a soft finger; For the three soft fingers in the first The average value of the predicted contact force for the predicted step; For the first The predicted circumferential torque proxy of the prediction step; The target circumferential torque proxy; This is the mode switching coefficient, which is 0 in grasp mode and 1 in twist mode; , , , , All are non-negative weights; To control the increment; For the first The soft finger in the first Pressure increment of the side-swing chamber assembly during the control cycle; For the first The soft finger in the first Control the inward bending cavity pressure increment during the control cycle; The positive part operator is used to set negative values ​​to zero; each candidate pressure sequence satisfies the upper limit of cavity pressure, the upper limit of cavity pressure change rate, and the upper limit of finger contact force. The controller only executes the first pressure increment of the candidate sequence with the minimum total cost and reconstructs the candidate sequence in the next control cycle.

7. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 6, characterized in that, In step S4, a side-swing cycle is formed according to the phase-staggered timing sequence. This side-swing cycle is divided into three phase time periods, and each phase time period mainly activates one unidirectional side-swing cavity group; The soft finger in the first The upper limit of the candidate pressure of the side-swing chamber during the control cycle is determined by the following formula: in, For the first The soft finger in the first Upper limit of candidate pressure for the side-swing chamber during the control cycle. The rated pressure of the side-swing chamber assembly. For the first The soft finger in the first Controlling the side-swing chamber pressure during the control cycle For the first The minimum holding force of a soft finger, This is the maximum allowable inward compensation pressure increment for this control cycle. This is a regularization term.

8. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 7, characterized in that, In step S5, four types of instantaneous residual sources are defined: contact force drop channel, contact force prediction channel, slip channel, and base reaction angular velocity channel. Specifically, the contact force drop channel takes the maximum value of the contact force decrease over three adjacent control cycles of the soft fingers; the contact force prediction channel takes the maximum value of the difference between the predicted and measured contact forces of the three soft fingers in one step; the slip channel takes the maximum value of the slip increment over three adjacent control cycles of the three soft fingers; and the base reaction angular velocity channel takes the around-axis reaction angular velocity characteristic output by the inertial measurement unit. The instantaneous residuals are first unified into dimensionless non-negative residuals according to the following formula: In the formula, For channel indexing, , , , , These correspond to the contact force drop channel, the base reaction angular velocity channel, the contact force prediction channel, and the sliding channel, respectively. For the first The channel is in The dimensionless nonnegative residual of the control period; Direction coefficient; For the first The channel is in The instantaneous residual value of the control cycle; For the first The average operating baseline of the channel; For the first The operating baseline variance of the channel; For regularization; the mean and variance of the operating baseline are determined within the non-separation steady-state window after entering the twist-and-pinch mode, and frozen within a single separation decision window; Then update the bounded continuous cumulative amount according to the following formula: In the formula, For the first The channel is in Bounded cumulative amount over the control period; For the first The channel is in Bounded cumulative amount over the control period; For memory coefficient; The term represents the non-negative drift term; and the separation risk is obtained by fusing the terms according to the following formula: In the formula, For the first Separation risk quantity during the control cycle; It is a collection of channels, including the contact force drop channel, the base reaction angular velocity channel, the contact force prediction channel, and the sliding channel; For the first Single-channel non-negative weights of the channel; For the first Channel and the Channel pairs have cooperative non-negative weights. For synergistic gain; For the first The channel is in The bounded cumulative amount of the control cycle; the stem separation judgment is output only when the separation risk amount reaches the separation threshold for several consecutive control cycles, and the bounded cumulative amount of at least two channels reaches the corresponding channel threshold.

9. The cross-mode cooperative control method for the multi-cavity pneumatic soft gripper of the harvesting robot according to claim 8, characterized in that, In step S6, the finger-by-finger target grasping force determined in step S3 is used as the holding reference, and the holding pressure is determined based on the real-time finger-by-finger contact force, inward bending cavity pressure, and finger-by-finger safe grasping force when the fruit stem is separated from the output fruit, so that the contact force of each soft finger is kept within the target holding range during the robot's movement; the release tolerance is determined by the center position of the basket opening, the normal direction of the basket opening, the outer size of the target fruit, and the preset safety gap, and after entering the release tolerance, the target fruit is released in the order of first stopping the side swing and then reducing the inward bending holding pressure.

Images