Soft grabbing arm for retail robot and control method of soft grabbing arm

Abstract

The invention provides a soft grabbing arm for a retail robot and a control method of the soft grabbing arm. The method belongs to the crossing field of soft robots, intelligent retail equipment and precise operation control. The method comprises the following steps: carrying out multi-cavity structure design on a software grabbing arm to generate multi-cavity software structure data; a distributed pressure sensor and micro-displacement sensor array is deployed according to the multi-cavity software structure data, and a multi-modal sensing network is constructed; through the combination of a multi-cavity soft structure and a multi-mode sensing network, data can be sensed online according to physical attributes of commodities, the pressure of each cavity can be dynamically adjusted, form self-adaptive grabbing is achieved, and the grabbing success rate of the vulnerable and special-shaped commodities is greatly increased.

Description

Technical Field

[0001] This invention proposes a soft gripping arm for retail robots and its control method, belonging to the interdisciplinary field of soft robots, intelligent retail equipment and precision operation control. Background Technology

[0002] In the field of intelligent retail logistics, robot-assisted product handling is a key step in improving efficiency and reducing labor costs. However, current retail robot handling systems have many problems and cannot meet the requirements for safe, damage-free, and high-precision handling of fragile and irregularly shaped goods.

[0003] In existing technologies, rigid grippers dominate, with two- or three-finger metal or plastic grippers operating based on preset clamping forces. This method lacks flexibility, and when dealing with easily deformable goods such as tomatoes and cakes, improper clamping force can easily crush or flatten them, causing damage to the goods.

[0004] While the simple pneumatic soft hand has been improved by using a silicone air bladder, it only controls the inflation of a single chamber, resulting in a fixed grasping shape. It struggles to effectively enclose irregularly shaped objects such as root vegetables, leading to poor grasping performance.

[0005] Moreover, most grasping systems use open-loop control, lacking a real-time force and deformation feedback mechanism during the grasping process. If the product slips or deforms beyond the preset range, the system cannot adjust dynamically in time, leading to grasping failure or product damage.

[0006] Furthermore, weak perception capabilities are a major drawback. Relying solely on an RGB camera to identify the product's position fails to capture key physical properties such as surface hardness and elastic modulus, hindering accurate grasping. Therefore, developing a novel soft gripper control method with adaptive shape, precise force control, multimodal perception, and closed-loop feedback capabilities is urgently needed. Summary of the Invention

[0007] This invention provides a soft gripping arm for retail robots and a control method thereof, to solve the problems mentioned in the background art above:

[0008] This invention proposes a control method for a soft gripper arm used in retail robots, the method comprising:

[0009] S1. Design a multi-chamber structure for the soft gripper arm and generate multi-chamber soft structure data; deploy a distributed pressure sensor and micro-displacement sensor array based on the multi-chamber soft structure data to construct a multimodal sensing network;

[0010] S2. Based on the multimodal sensing network, pressure distribution data and surface deformation data of the contact area of ​​the goods are collected in real time to generate online sensing data of the physical properties of the goods, and drive the pneumatic inverse model planning module to generate the optimal pressure control command for each chamber.

[0011] S3. Implement multi-chamber coordinated inflation / deflation operation according to the optimal pressure control command, generate dynamic deformation adjustment data, drive the end effector of the soft gripper arm to perform shape adaptive adjustment, and generate envelope gripping posture data that perfectly fits the shape of the product.

[0012] S4. Utilize a distributed pressure sensor array to monitor changes in contact force during the grasping process in real time and generate tactile feedback data; simultaneously acquire spatial pose information of the product through an RGB-D camera to generate visual feedback data; fuse the tactile feedback data and visual feedback data into a dual-modal closed-loop control signal;

[0013] S5. Dynamically correct the output parameters of the pneumatic inverse model planning module based on the dual-mode closed-loop control signal to generate real-time optimized control commands; drive the soft gripper arm to perform precise force control through the real-time optimized control commands to generate non-destructive gripping execution data.

[0014] S6. Based on the non-destructive gripping execution data, calculate the deformation rate and contact stress distribution uniformity index of the goods, generate gripping quality assessment data, and dynamically adjust the sampling frequency of the multimodal sensing network and the iteration cycle of the aerodynamic inverse model planning module to generate adaptive control parameters, which are fed back to the multi-chamber software structure data layer to form a full-process closed-loop control system and generate the goods gripping control data stream.

[0015] This invention proposes a soft gripper arm for retail robots, comprising:

[0016] One or more processors;

[0017] Memory, used to store one or more programs;

[0018] Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0019] The beneficial effects of this invention are as follows: By combining a multi-chamber soft structure with a multimodal sensing network, the pressure in each chamber can be dynamically adjusted based on online sensing data of the physical properties of the goods, achieving shape-adaptive gripping and greatly improving the success rate of gripping fragile and irregularly shaped goods. Simultaneously, the synergistic effect of pneumatic inverse model planning and dual-modal closed-loop control can precisely regulate the gripping force, reducing the probability of goods being damaged due to improper gripping force and effectively avoiding economic losses. Moreover, the rich data collected in real time by the multimodal sensing network enhances the system's ability to perceive key attributes such as the surface hardness and elastic modulus of the goods, making the gripping process more precise. Furthermore, the full-process closed-loop control system reduces the need for manual intervention, lowering labor costs. This control method can adapt to gripping goods of different shapes and textures and can dynamically optimize control parameters based on real-time feedback, avoiding the limitations of traditional rigid grippers or simple pneumatic soft grippers, providing a more efficient and reliable flexible operation solution for intelligent retail logistics. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the steps of the method described in this invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] One embodiment of the present invention, such as Figure 1 As shown, a control method for a soft gripper arm used in a retail robot includes:

[0023] S1. Design a multi-chamber structure for the soft gripper arm and generate multi-chamber soft structure data; deploy a distributed pressure sensor and micro-displacement sensor array based on the multi-chamber soft structure data to construct a multimodal sensing network;

[0024] S2. Based on a multimodal sensing network, real-time pressure distribution data and surface deformation data of the product contact area are collected to generate online sensing data of the product's physical properties; the online sensing data of the product's physical properties drives the pneumatic inverse model planning module to generate the optimal pressure control command for each chamber;

[0025] S3. Implement multi-chamber coordinated inflation / deflation operation according to the optimal pressure control command to generate dynamic deformation adjustment data; drive the end effector of the soft gripping arm to perform shape adaptive adjustment through the dynamic deformation adjustment data to generate envelope gripping posture data that perfectly fits the shape of the product.

[0026] S4. Utilize a distributed pressure sensor array to monitor changes in contact force during the grasping process in real time and generate tactile feedback data; simultaneously acquire spatial pose information of the product through an RGB-D camera to generate visual feedback data; fuse the tactile feedback data and visual feedback data into a dual-modal closed-loop control signal;

[0027] S5. Dynamically correct the output parameters of the pneumatic inverse model planning module based on the dual-mode closed-loop control signal to generate real-time optimized control commands; drive the soft gripper arm to perform precise force control through the real-time optimized control commands to generate non-destructive gripping execution data.

[0028] S6. Calculate the deformation rate and contact stress distribution uniformity index of the goods based on the non-destructive gripping execution data to generate gripping quality assessment data; dynamically adjust the sampling frequency of the multimodal sensing network and the iteration cycle of the aerodynamic inverse model planning module according to the gripping quality assessment data to generate adaptive control parameters; feed the adaptive control parameters back to the multi-chamber software structure data layer to form a full-process closed-loop control system and generate a safe, non-destructive, and high-precision goods gripping control data stream.

[0029] The working principle and effects of the above technical solution are as follows: Through the collaboration of a multimodal perception network and a pneumatic inverse model planning module, the physical attributes of goods are accurately captured, significantly improving the posture adaptation accuracy of the soft gripping arm and effectively avoiding damage to fragile goods caused by improper gripping force; the dual-modal closed-loop control integrates tactile and visual feedback in real time, dynamically corrects control parameters, enhances the stability of the gripping process, and reduces errors caused by contact force fluctuations and posture deviations; adaptive adjustment of the perception sampling frequency and model iteration cycle reduces the system's ineffective energy consumption and flexibly adapts to goods of different sizes, weights, and fragility levels, avoiding the limitations of insufficient adaptability of a single control parameter; the gripping quality assessment and full-process feedback mechanism ensures damage-free gripping of goods and continuously optimizes the control logic, significantly improving the gripping reliability in retail scenarios and reducing losses during goods handling; the overall solution balances gripping accuracy and operational efficiency, making retail robots more practical in complex goods gripping scenarios and effectively improving the problems of traditional gripping methods that easily damage goods and have poor adaptability.

[0030] In one embodiment of the present invention, S1 includes:

[0031] S11. Investigate the size range (5cm-50cm), weight range (0.1kg-5kg), and fragility level (1-5) of typical products in retail scenarios, and generate scenario adaptability design parameters in combination with robot operating space constraints.

[0032] S12. Based on the design parameters, optimize the multi-chamber topology to determine the number of chambers (4-8), the annular distribution density, and the wall thickness gradient (0.5mm-2mm), and generate initial data for the multi-chamber soft structure.

[0033] S13. Perform ANSYS fluid dynamics simulation and physical prototype deformation characteristic test on the initial data, correct the chamber layout and the Shore hardness parameters of the silicone material, and generate the final data of the multi-chamber soft structure.

[0034] S14. Based on the final data, plan the sensor deployment sites (avoiding the critical area of ​​chamber inflation and expansion), and deploy a distributed pressure sensor (range 0-10N) and micro-displacement sensor (accuracy 0.01mm) array at a spacing of 3cm×3cm;

[0035] S15. Perform zero-point calibration, 50Hz low-pass filtering, and Modbus communication protocol adaptation on the sensor array to construct a multimodal sensing network.

[0036] The working principle and effects of the above technical solution are as follows:

[0037] In one embodiment of the present invention, S13 includes:

[0038] Extract the geometric dimensions, chamber spacing, and initial material hardness parameters from the initial data of the multi-chamber soft structure, construct a 1:1 ANSYS fluid dynamics simulation model, and generate the basic simulation modeling file;

[0039] Set the simulation boundary conditions (inflation pressure range 0.1-0.8MPa, fixed constraint points) and solution accuracy parameters, run the fluid and structure coupled simulation, output the chamber deformation trajectory, stress concentration area, pressure transmission efficiency data, and generate a simulation analysis result set;

[0040] Based on the initial data, a physical prototype was made, and a deformation characteristic testing platform (including a pressure loading module and a laser three-dimensional scanning device) was built. The prototype was subjected to graded pressure loading tests, and the actual deformation amplitude and stress distribution hotspot data were collected to generate a physical test dataset.

[0041] By comparing the simulation results set with the physical test dataset, the deformation error (allowable threshold ≤5%) and stress deviation are calculated, the congested areas of the chamber layout and the material hardness mismatch problem are identified, and a parameter correction priority list is generated.

[0042] Adjust the chamber distribution density and optimize the local wall thickness according to the priority list, iteratively correct the Shore hardness parameters of the silicone material (adjustment range 10-30HA), and perform simulation verification again until the error meets the requirements, and generate the final data of the multi-chamber soft structure.

[0043] The working principle and effects of the above technical solution are as follows: Through bidirectional verification of simulation modeling and physical testing, the deformation characteristics and stress distribution of the multi-chamber structure are accurately captured, significantly improving the accuracy of the software structure design parameters; iterative correction of chamber layout, wall thickness, and material hardness parameters strictly controls the deformation error within 5%, effectively avoiding problems such as loss of gripping posture and stress concentration damage caused by structural design defects; it reduces the cost of repeated manufacturing and debugging of physical prototypes and shortens the structural development cycle, making the design solution more in line with the actual needs of retail scenarios. The sensor deployment sites avoid key deformation areas, and with the calibration and filtering of the sensor array, the stability and accuracy of the perceived data are enhanced, avoiding signal interference or data distortion from affecting subsequent control decisions; the overall process makes the multimodal sensing network highly compatible with the multi-chamber structure, laying a solid foundation for subsequent accurate identification of product attributes and non-destructive gripping, significantly improving the reliability and implementability of the entire control solution.

[0044] In one embodiment of the present invention, S2 includes:

[0045] S21. Activate the real-time acquisition unit of the multimodal sensing network to pre-scan and sample the potential contact area of ​​the product at a sampling frequency of 100Hz, and generate the original sensing data of the contact area.

[0046] S22. The Kalman filter algorithm is used to remove electromagnetic interference noise from the raw sensing data, and combined with Z-score standardization, a clean stress and deformation dataset is generated.

[0047] S23. Extract 12-dimensional feature parameters (such as peak contact force, deformation gradient, surface roughness, etc.) from the dataset, reduce the dimensionality to 6 dimensions through principal component analysis, and generate online perception data of commodity physical properties.

[0048] S24. Input the sensing data into the feature matching layer of the pneumatic inverse model planning module, and establish a mapping relationship model between commodity attributes and chamber pressure based on the BP neural network;

[0049] S25. Solve the optimal solution of the mapping model using the particle swarm optimization algorithm to generate the initial pressure control command for each chamber. After ±5% disturbance stability verification, correct the command amplitude and timing parameters to generate the optimal pressure control command.

[0050] The working principle and effects of the above technical solution are as follows: By combining 100Hz high-frequency sampling with Kalman filtering and standardization, the cleanliness of the sensing data in the contact area is significantly improved, effectively reducing noise pollution caused by electromagnetic interference and avoiding misjudgment of product attributes due to distortion of the original data; dimensionality reduction and extraction of 6 core parameters from 12-dimensional features not only fully preserves the key information of the product's physical attributes but also reduces the computational load of the model, significantly improving data processing efficiency; a mapping relationship between attributes and chamber pressure is established based on a BP neural network, and the optimal solution is obtained by combining it with a particle swarm optimization algorithm, allowing the pressure control command to be accurately matched with the product characteristics, improving the scientific nature of command generation; ±5% disturbance stability verification further corrects the parameters, enhancing the anti-interference capability of the control command and avoiding grasping deviation caused by command failure under sudden working conditions; the entire process realizes full-link optimization from data acquisition and processing to command generation, ensuring the accuracy of product physical attribute identification and improving the reliability and stability of pressure control commands, providing precise guidance for subsequent multi-chamber collaborative deformation and precise envelope grasping, and effectively reducing the risk of product damage or grasping failure caused by improper commands.

[0051] In one embodiment of the present invention, S24 includes:

[0052] Extract the 6-dimensional core feature vector from the online sensing data of product physical attributes, label the attribute tags according to data type (such as weight category, deformation category), and generate a standardized input dataset for BP neural network;

[0053] Based on the input dimension (6 dimensions) and the output dimension (the number of chambers corresponds to the pressure dimension), a 3-layer topology of a BP neural network (input layer - hidden layer - output layer) is designed. The number of neurons in the hidden layer (2-3 times the input dimension), the ReLU activation function and the mean squared error loss function are set to generate network structure configuration parameters.

[0054] The system retrieves product attribute and chamber pressure matching samples from the historical database of the retail robot, merges them with the current input dataset, divides the training set and validation set in a 7:3 ratio, and generates a model training sample set.

[0055] Input the training sample set into the configured BP neural network, start iterative training (set the number of iterations to 500 and the learning rate to 0.001), dynamically update the network weights and biases through the backpropagation algorithm, record the training error curve in real time, and generate a dataset of the model training process.

[0056] The model's prediction accuracy is verified using a validation set. The relative error between the predicted and actual pressure values ​​is calculated (with an allowable threshold of ≤3%). If the error exceeds the threshold, the learning rate and the number of neurons in the hidden layer are adjusted and the model is retrained. Once the threshold is met, a mapping relationship model between product attributes and chamber pressure is generated.

[0057] The working principle and effects of the above technical solution are as follows: By labeling the 6-dimensional core feature vector with attribute tags and standardizing the processing, the model input bias caused by data type confusion is effectively avoided, and the adaptability of data and network structure is improved; the rational design of the 3-layer topology, combined with the precise configuration of the number of hidden layer neurons, activation function and loss function, enhances the model's ability to fit the mapping relationship between commodity attributes and chamber pressure, and avoids the problem of inefficient prediction caused by network structure redundancy or insufficiency; by integrating historical crawled samples with the current dataset and dividing the training and validation sets in a 7:3 ratio, the model's generalization ability can be fully utilized to improve its generalization ability, and the prediction accuracy can be accurately verified through the validation set, reducing the risk of overfitting caused by training on a single dataset. The dynamic parameter updates of the backpropagation algorithm, with 500 iterations of training and adaptive adjustments after the error exceeds the limit, strictly control the relative error of pressure prediction within 3%, which greatly improves the prediction accuracy of the mapping model. The whole process makes the mapping relationship between product attributes and chamber pressure more accurate, avoids the inaccuracy of pressure control commands caused by model prediction deviations, provides reliable support for the generation of subsequent optimal pressure commands, reduces the blindness of model debugging, and reduces the potential risk of product damage or grasping failure during subsequent grasping process, making the control scheme more practical and stable.

[0058] In one embodiment of the present invention, S3 includes:

[0059] S31. Parse the optimal pressure control command and convert it into the inflation / deflation flow rate threshold (0-5L / min), pressure target value (0.1-0.8MPa), and timing logic parameters for each chamber, and generate the chamber pneumatic control execution parameter set;

[0060] S32. Based on the parameter set, the pneumatic proportional valve group is driven to perform multi-chamber coordinated inflation / deflation operation, and the chamber pressure feedback data is collected in real time through the pressure transmitter (sampling frequency 50Hz).

[0061] S33. By fusing pressure feedback data with deformation monitoring data from the sensor array, a weighted fusion algorithm is used to generate dynamic deformation adjustment data;

[0062] S34. Input the adjustment data into the end effector shape control algorithm based on Bézier curves to generate segmented attitude adjustment path (coarse positioning segment + fine adaptation segment) data.

[0063] S35. Based on the path data, the actuator is adjusted, and the fit is verified in real time by an RGB-D camera (error ≤ 0.5mm). The parameters are corrected to generate envelope-type grasping posture data that perfectly fits the shape of the product.

[0064] The working principle and effects of the above technical solution are as follows: It parses the optimal pressure control command and transforms it into a specific set of execution parameters, clearly defining the flow rate, pressure, and timing logic of each chamber, effectively avoiding operational confusion caused by improper adaptation between the command and the pneumatic actuator, and improving the implementation of control commands; based on the parameter set, it drives multi-chamber coordinated inflation / deflation, coupled with 50Hz high-frequency pressure feedback acquisition, enhancing the real-time performance and accuracy of chamber pressure regulation, and reducing deformation deviation caused by pressure overshoot or lag; through a weighted fusion algorithm, it integrates pressure feedback and deformation monitoring data, making the dynamic deformation adjustment data more closely match the actual operating conditions, avoiding the limitations of a single data source. The system avoids adjustment misalignment caused by inaccuracies; the segmented attitude adjustment path driven by Bézier curves can improve adjustment efficiency through coarse positioning segments and ensure fitting accuracy through fine fitting segments, balancing operation speed and control quality; the RGB-D camera verifies the fitting degree and corrects parameters in real time, strictly controlling the error within 0.5mm, effectively avoiding problems such as unstable gripping or excessive local force on the product caused by inadequate attitude adjustment, and significantly improving the fitting accuracy and reliability of envelope gripping; the entire process makes the shape adjustment of the soft gripping arm more targeted, laying a solid attitude foundation for subsequent non-destructive gripping, while reducing ineffective adjustment actions and lowering energy consumption and mechanical wear.

[0065] In one embodiment of the present invention, step S4 includes:

[0066] S41. Start the distributed pressure sensor array at a 50Hz monitoring frequency to continuously collect dynamic change data of contact force during the grasping process and generate raw tactile feedback data stream.

[0067] S42. Perform peak detection, moving average filtering and feature extraction on the raw data stream to generate standardized tactile feedback data containing the average contact force and fluctuation amplitude.

[0068] S43. Start the RGB-D camera and aim it at the work area. Collect the spatial position coordinates, attitude angle (Roll / Pitch / Yaw) and displacement change information of the product at a frame rate of 30fps to generate raw visual feedback data.

[0069] S44. Perform distortion correction and robot base coordinate system calibration on the original visual data, and use the ICP algorithm to unify the coordinate system and generate standardized visual feedback data.

[0070] S45. Construct a weighted average fusion model, input dual-modal standardized data, and generate a high-precision dual-modal closed-loop control signal after eliminating redundant interference.

[0071] The working principle and effects of the above technical solution are as follows: 50Hz high-frequency tactile acquisition combined with peak detection and moving average filtering can capture subtle fluctuations in contact force in real time, effectively reducing noise pollution caused by electromagnetic interference and mechanical vibration, making standardized tactile data more realistic and reliable, and avoiding control inaccuracies caused by force feedback lag. RGB-D camera acquisition at 30fps, combined with distortion correction and ICP algorithm coordinate system one, significantly improves the accuracy of product pose information, eliminates judgment errors caused by different coordinate system deviations, and accurately reflects product displacement and posture changes; the weighted average fusion model integrates dual-modal data, leveraging the sensitivity of tactile perception to contact force while utilizing the accurate judgment of pose by vision, effectively eliminating data redundancy and cross-interference, generating a high-precision closed-loop control signal that provides a reliable basis for subsequent parameter correction; the entire process enhances the comprehensiveness and accuracy of grasping feedback, avoiding posture deviation or uneven force caused by one-sided single-modal feedback, significantly improving the response speed and stability of closed-loop control, reducing the risk of damage during product grasping, and making subsequent force control adjustments more targeted.

[0072] In one embodiment of the present invention, step S5 includes:

[0073] S51. Analyze the dual-mode closed-loop control signal, extract the contact force deviation value (≤±0.2N), attitude offset (≤±1°) and deformation error parameters, and generate control correction requirement data;

[0074] S52. Input the corrected demand data into the parameter correction layer of the aerodynamic inverse model planning module to dynamically adjust the model weight coefficients and solution boundary conditions.

[0075] S53. Based on the corrected model, recalculate iteratively to generate preliminary optimized control commands, and then correct the parameters after force control accuracy verification (error ≤ 5%).

[0076] S54. Send real-time optimized control commands to the pneumatic actuator to drive dynamic fine-tuning of chamber pressure, simultaneously monitor contact force stability and product deformation status, and generate non-destructive gripping execution data.

[0077] The working principle and effect of the above technical solution are as follows: it analyzes the dual-mode closed-loop control signal, accurately extracts core parameters such as contact force deviation (≤±0.2N) and attitude offset (≤±1°), clearly defines the core direction of control correction, avoids the waste of resources caused by blind adjustment, and improves the pertinence and effectiveness of correction requirement data. The corrected data is input into the parameter correction layer of the pneumatic inverse model, dynamically adjusting the weight coefficients and solution boundaries. This allows the model to adapt to the dynamic changes in the grasping conditions in real time, enhancing its adaptability and reducing the limitations of fixed-parameter models in adapting to complex scenarios. Based on the corrected model, iterative calculations are performed, coupled with force control accuracy verification of ≤5%, effectively eliminating unqualified initial commands and significantly improving the accuracy of control commands. This avoids problems such as excessive contact force and loss of posture caused by command errors. Real-time optimization commands drive dynamic fine-tuning of chamber pressure, simultaneously monitoring contact force stability and product deformation status. This allows for rapid correction of minor deviations during the grasping process while strictly controlling the product's safety margin, resulting in more reliable, damage-free grasping execution data. The entire process makes closed-loop control more responsive and accurate, significantly reducing the risk of damage to fragile products during grasping. It also reduces the time cost of repeated control parameter adjustments, making the retail robot's grasping operation more stable and efficient. Furthermore, it provides high-quality data support for subsequent grasping quality assessment and adaptive parameter optimization, further solidifying the reliability of the entire closed-loop control process.

[0078] In one embodiment of the present invention, S54 includes:

[0079] The real-time optimization control commands are analyzed, and the pressure regulation amplitude, fine-tuning rate and timing synchronization parameters of each chamber are extracted and converted into an electrical signal format adapted to the pneumatic actuator to generate a set of chamber pressure fine-tuning execution signals.

[0080] The execution signal set is sent to the pneumatic proportional valve group and pressure regulation module to drive each chamber to perform dynamic pressure fine-tuning according to preset parameters, and the chamber pressure feedback acquisition unit (sampling frequency 100Hz) is started simultaneously to generate real-time pressure feedback data stream.

[0081] The system calls upon a distributed pressure sensor array to monitor dynamic fluctuations in contact force, combines this with real-time deformation data of the product collected by a micro-displacement sensor, and merges the two types of data to generate a dual-dimensional monitoring data stream of contact force and deformation.

[0082] Analyze the monitoring data stream to determine whether the contact force fluctuation amplitude (allowable threshold ≤ ±0.1N) and the product deformation value (allowable threshold ≤ 0.3mm) meet the non-destructive requirements. If they exceed the threshold, generate an instant correction signal to dynamically adjust the fine-tuning rate and pressure amplitude.

[0083] Once the pressure stabilizes and the monitoring data meets the non-destructive standards, the fine-tuning operation is stopped. The pressure adjustment parameters, contact force stability data, and product deformation compliance data are integrated to generate non-destructive gripping execution data.

[0084] The working principle and effects of the above technical solution are as follows: Real-time optimization control commands are parsed and converted into electrical signals adapted to the pneumatic actuator, effectively avoiding execution delays or malfunctions caused by incompatible signal formats, and improving the efficiency of pressure fine-tuning commands; 100Hz high-frequency pressure feedback acquisition can accurately capture minute fluctuations in chamber pressure, enhancing the real-time responsiveness of the fine-tuning process and reducing control deviations caused by pressure regulation lag; dual-dimensional monitoring of dynamic fluctuations in contact force and real-time deformation of the product not only comprehensively grasps the grasping status but also, through contact force fluctuation thresholds of ≤±0.1N and ≤0 Strict control of the 0.3mm deformation threshold effectively prevents damage to goods caused by uneven force or excessive deformation. The instant correction mechanism when the threshold is exceeded dynamically adjusts the fine-tuning rate and pressure amplitude, making pressure adjustment more flexible and further solidifying the bottom line of non-destructive grasping. The non-destructive grasping execution data generated after the data meets the standards not only ensures data integrity and reliability, but also provides an accurate basis for subsequent grasping quality assessment. At the same time, it reduces the loss rate of goods grasping in retail scenarios, making the operation of the soft grasping arm safer and more controllable, and also improving the adaptability of grasping complex goods.

[0085] In one embodiment of the present invention, step S6 includes:

[0086] S61. Extract the three-dimensional contour point cloud data and contact stress acquisition data of the product before and after grasping from the non-destructive grasping execution data to generate the original dataset for quality assessment.

[0087] S62. Calculate the deformation (profile deviation value) and deformation rate (deformation / original size) of the commodity, analyze the standard deviation of contact stress distribution through finite element interpolation, and generate deformation and stress uniformity indexes.

[0088] S63. Set the weights of the indicators (deformation rate 0.6, stress uniformity 0.4), calculate the comprehensive grasping quality score (≥85 points is qualified), and generate grasping quality assessment data;

[0089] S64. Dynamically adjust the sampling frequency of the sensing network (80Hz for qualified / 120Hz for unqualified) and the model iteration period (50ms for qualified / 30ms for unqualified) based on the evaluation data to generate an adaptive control parameter set;

[0090] S65. Feed the parameter set back to the multi-chamber software structure data layer to build a closed-loop control system for the entire process of perception, planning, execution, evaluation and feedback, and generate a safe, non-destructive and high-precision commodity grasping control data stream.

[0091] The working principle and effects of the above technical solution are as follows: Three-dimensional contour point clouds and contact stress information are extracted from the non-destructive grasping execution data to generate a raw dataset for quality assessment, providing solid data support for the assessment and avoiding biases caused by subjective judgment; by calculating the deformation rate and analyzing the standard deviation of stress distribution, coupled with scientific indicator weighting and a passing score standard, the objectivity and accuracy of grasping quality assessment are significantly improved, clearly defining the quality of grasping performance and reducing the problem of unclear optimization direction caused by fuzzy assessment; the sensing sampling frequency and model iteration cycle are dynamically adjusted according to the assessment results. When the performance is qualified, 80Hz sampling and a 50ms iteration cycle are used to reduce system energy consumption; when the performance is unqualified, the sampling frequency and model iteration cycle are adjusted accordingly. Upgrading to 120Hz and 30ms ensures accuracy, avoiding resource waste and specifically addressing shortcomings. Feeding adaptive parameters back to the structured data layer creates a closed-loop control system, enhancing the system's self-optimization and adaptability. This prevents the decline in grasping quality caused by fixed parameters over long-term use, reducing the frequency and cost of manual debugging. The entire process creates a virtuous cycle for the retail robot's grasping control, continuously improving the stability of safe and lossless grasping while allowing the system to flexibly respond to the grasping needs of different products, further reducing product loss rates. This makes the control solution more long-lasting and practical, providing reliable assurance for efficient operation in complex retail scenarios.

[0092] In one embodiment of the present invention, S64 includes:

[0093] The data is analyzed to capture quality assessment data, extract key indicators such as comprehensive capture quality score, deformation rate deviation from threshold, and stress uniformity coefficient, and generate a dataset for parameter adjustment decision-making.

[0094] Based on the scoring results, the pass / fail status is determined (≥85 points is pass). The sampling frequency of the sensing network is initially determined according to the basic rules (80Hz for pass / 120Hz for fail). The frequency value is then finely adjusted based on the degree of deformation rate deviation (the basic value is maintained if the deviation is ≤10%, and ±10Hz is maintained if the deviation is >10%) to generate candidate sampling frequency values.

[0095] Similarly, determine the initial value of the model iteration cycle according to the qualified state (50ms for qualified / 30ms for unqualified), optimize the cycle parameters according to the stress uniformity coefficient (keep the initial value if the coefficient is ≥0.8, shorten it by 5-10ms if <0.8), and generate candidate values ​​for the iteration cycle;

[0096] Perform a collaborative compatibility check on the candidate values ​​of sampling frequency and iteration period (to avoid system overload caused by high frequency and short period), calculate the system load rate corresponding to the parameter combination (allowing threshold ≤70%), correct incompatible combinations, and generate parameter combination schemes;

[0097] After verifying the control response effect of the parameter combination scheme and confirming that the scheme can improve the grasping accuracy or reduce the system energy consumption, the frequency and period parameters are integrated to generate an adaptive control parameter set.

[0098] The working principle and effects of the above technical solution are as follows: Analyzing and capturing quality assessment data and extracting key indicators provides a clear basis for parameter adjustments, avoiding resource waste caused by blind adjustments and significantly improving the targeting of parameter optimization; determining basic parameters based on the qualified state, and then combining the degree of deformation rate deviation and stress uniformity coefficient for precise fine-tuning, can reduce system energy consumption with a lower sampling frequency (80Hz) and a longer iteration period (50ms) when the data is qualified, while improving parameter assurance accuracy when the data is unqualified, thus balancing control effect and resource efficiency; collaborative compatibility verification strictly controls the system load rate (≤70%). This effectively avoids system overload and response lag caused by the superposition of high frequency and short cycle, and enhances the stability and feasibility of parameter combinations. The solution verification stage ensures that parameter adjustments can effectively improve grasping accuracy or reduce energy consumption, reducing the probability of invalid adjustments and making the generated adaptive control parameter set more practical. The entire process makes the parameter adjustment of the perception network and model more scientific and flexible, avoiding the rigid limitations of single parameter rules and accurately adapting to different grasping quality scenarios. It provides reliable parameter support for the whole process closed-loop control and further improves the adaptive capability and long-term stability of the retail robot's grasping control.

[0099] One embodiment of the present invention provides a soft gripper arm for a retail robot, comprising:

[0100] One or more processors;

[0101] Memory, used to store one or more programs;

[0102] Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0103] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A control method for a soft gripper arm of a retail robot, characterized in that, The method comprises: S1, a multi-chamber structure design is performed on the soft body grabbing arm, multi-chamber soft structure data is generated, a distributed pressure sensor and a micro-displacement sensor array are deployed according to the multi-chamber soft structure data, and a multi-modal perception network is constructed; S2, based on the multi-modal perception network, pressure distribution data and surface deformation data of a commodity contact area are collected in real time, commodity physical property online perception data are generated, a pneumatic inverse model planning module is driven, and optimal pressure control instructions of each chamber are generated; S3, according to the optimal pressure control instructions, a multi-chamber coordinated inflation / deflation operation is implemented, dynamic deformation adjustment data are generated, a form self-adaptive adjustment of the end effector of the soft body grabbing arm is driven, and envelope type grabbing posture data completely conforming to the commodity shape are generated; S4, the distributed pressure sensor array is used to monitor the contact force change in the grabbing process in real time, touch feedback data are generated; the spatial pose information of the commodity is collected synchronously through the RGB-D camera, visual feedback data are generated; the touch feedback data and the visual feedback data are fused into a dual-modal closed-loop control signal; S5, according to the dual-modal closed-loop control signal, the output parameters of the pneumatic inverse model planning module are dynamically corrected, real-time optimization control instructions are generated; the soft body grabbing arm is driven to implement force precise regulation and control through the real-time optimization control instructions, and lossless grabbing execution data are generated; S6, based on the lossless grabbing execution data, commodity deformation rate and contact stress distribution uniformity indexes are calculated, grabbing quality evaluation data are generated, the sampling frequency of the multi-modal perception network and the iteration period of the pneumatic inverse model planning module are dynamically adjusted, adaptive control parameters are generated, and the adaptive control parameters are fed back to the multi-chamber soft structure data layer to form a whole-process closed-loop control system, and commodity grabbing control data flow is generated.

2. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S1 comprises: S11, investigating the size interval, weight range and damageable level of typical commodities in a retail scene, combining with the robot operation space constraint, and generating scene adaptability design parameters; S12, performing multi-chamber topological structure optimization based on the design parameters, determining the number of chambers, annular distribution density and wall thickness gradient, and generating multi-chamber soft structure initial data; S13, performing ANSYS fluid mechanics simulation and physical prototype deformation characteristic test on the initial data, correcting the chamber layout and the Shore hardness parameters of the silica gel material, and generating multi-chamber soft structure final data; S14, planning sensor deployment sites according to the final data, and deploying a distributed pressure sensor and a micro-displacement sensor array at an interval of 3cm*3cm; S15, performing zero point calibration, 50Hz low pass filter processing and Modbus communication protocol adaptation on the sensor array, and constructing a multi-modal perception network.

3. The control method of a software gripper arm for a retail robot according to claim 2, wherein, The S13 comprises: geometric dimensions, chamber spacing and initial material hardness parameters in the multi-chamber soft structure initial data are extracted, a 1:1 ANSYS fluid mechanics simulation model is constructed, and a simulation modeling basis file is generated; simulation boundary conditions and solution accuracy parameters are set, fluid and structure coupling simulation is run, chamber deformation trajectory, stress concentration area and pressure conduction efficiency data are output, and a simulation analysis result set is generated; According to the initial data, a physical prototype is made, a deformation characteristic test platform is built, a graded pressure loading test is performed on the prototype, actual deformation amplitude, stress distribution hot spot data are collected, and a physical test data set is generated; Compare the simulation result set with the physical test data set, calculate the deformation error and stress deviation, identify the congestion area of the chamber layout and the mismatch of the material hardness, and generate a parameter correction priority list; According to the priority list, adjust the chamber distribution density and optimize the local wall thickness, iteratively correct the Shore hardness parameters of the silica gel material, perform simulation verification again until the error meets the requirements, and generate the final data of the multi-chamber soft structure.

4. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S2 comprises: S21, starting the real-time acquisition unit of the multi-modal perception network, pre-scanning and sampling the potential contact area of the commodity at a sampling frequency of 100Hz, and generating contact area original perception data; S22, using Kalman filtering algorithm to remove electromagnetic interference noise from the original perception data, and combining Z-score standardization processing to generate clean pressure and deformation data set; S23, extracting 12-dimensional feature parameters from the data set, reducing dimensionality to 6 dimensions through principal component analysis, and generating online perception data of commodity physical properties; S24, inputting the perception data into the feature matching layer of the pneumatic inverse model planning module, and establishing a mapping relationship model between commodity attributes and chamber pressure based on BP neural network; S25, solving the optimal solution of the mapping model through particle swarm optimization algorithm, generating initial pressure control instructions for each chamber, correcting the instruction amplitude and timing parameters after ±5% disturbance stability verification, and generating optimal pressure control instructions.

5. The control method of a software gripper arm for a retail robot according to claim 4, wherein, The S24 comprises: Extracting 6-dimensional core feature vectors in online perception data of commodity physical properties, labeling attribute tags according to data types, and generating BP neural network standardized input data set; According to the input dimension and output dimension, designing a 3-layer topology structure of BP neural network, setting the number of hidden layer neurons, ReLU activation function and mean square error loss function, and generating network structure configuration parameters; Calling commodity attributes and chamber pressure matching samples in the historical grabbing database of retail robots, fusing with the current input data set, dividing the training set and the verification set according to the 7:3 ratio, and generating model training sample set; Inputting the training sample set into the completed BP neural network, starting the iterative training, dynamically updating the network weight and bias through the back propagation algorithm, recording the training error curve in real time, and generating model training process data set; Verifying the prediction accuracy of the model with the verification set, calculating the relative error between the pressure predicted value and the actual value, adjusting the learning rate and the number of hidden layer neurons if the error is out of tolerance, and generating the mapping relationship model between commodity attributes and chamber pressure after reaching the standard.

6. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S3 comprises: S31, analyzing the optimal pressure control instructions, converting them into chamber inflation / deflation flow threshold, pressure target value and timing logic parameters, and generating chamber pneumatic control execution parameter set; S32, driving the pneumatic proportional valve group to perform multi-chamber collaborative inflation / deflation operation based on the parameter set, and collecting chamber pressure feedback data in real time through the pressure transmitter; S33, fuse the pressure feedback data with the deformation monitoring data of the sensor array, generate dynamic deformation adjustment data using a weighted fusion algorithm; S34, input the adjustment data into a end effector morphology control algorithm based on Bezier curve, generate segmented posture adjustment path data; S35, drive the effector adjustment according to the path data, real-time verify the fit degree through the RGB-D camera, correct the parameters to generate an envelope type grabbing posture data completely fitted with the product shape.

7. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S4 comprises: S41, start the distributed pressure sensor array at a monitoring frequency of 50Hz, continuously collect the contact force dynamic change data during the grabbing process, and generate the original tactile feedback data stream; S42, perform peak detection, sliding average filtering and feature extraction on the original data stream, and generate standardized tactile feedback data containing contact force mean and fluctuation amplitude; S43, start the RGB-D camera to align the work area, collect the spatial position coordinates, attitude angle and displacement change information of the product at a frame rate of 30fps, and generate original visual feedback data; S44, perform distortion correction processing and robot base coordinate system calibration on the original visual data, align the coordinate systems through ICP algorithm, and generate standardized visual feedback data; S45, construct a weighted average fusion model, input the double-mode standardized data, and generate high-precision double-mode closed-loop control signals after eliminating redundant interference.

8. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S5 comprises: S51, analyze the double-mode closed-loop control signal, extract the contact force deviation value, attitude offset and deformation error parameters, and generate control correction requirement data; S52, input the correction requirement data into the parameter correction layer of the pneumatic inverse model planning module, dynamically adjust the model weight coefficient and solve the boundary conditions; S53, based on the corrected model, reiterate calculation to generate preliminary optimized control instructions, and correct the parameters after force control accuracy verification; S54, send the real-time optimized control instructions to the pneumatic execution unit, drive the cavity pressure dynamic fine adjustment, synchronously monitor the contact force stability and the product deformation state, and generate non-destructive grabbing execution data.

9. The control method of a software gripper arm for a retail robot according to claim 1, wherein, The S6 comprises: S61, extract the three-dimensional contour point cloud data and contact stress collection data before and after the product grabbing from the non-destructive grabbing execution data, and generate a quality evaluation original data set; S62, calculate the product deformation and deformation rate, analyze the standard deviation of contact stress distribution through finite element interpolation method, and generate deformation and stress uniformity index; S63, set the index weight, calculate the comprehensive grabbing quality score, and generate the grabbing quality evaluation data; S64, dynamically adjust the sampling frequency and model iteration period of the perception network according to the evaluation data, and generate an adaptive control parameter set; S65, feedback the parameter set to the multi-cavity soft structure data layer, construct a perception, planning, execution, evaluation and feedback whole-process closed-loop control system, and generate safe, non-destructive and high-precision product grabbing control data stream.

10. A soft grabbing arm for a retail robot, comprising: one or more processors; memory for storing one or more programs; wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to carry out the method of any one of claims 1 to 9.

Images