A humanoid robot upper limb end adaptive grasping control method, system and storage medium

Abstract

The application discloses a humanoid robot upper limb end adaptive grasping control method and system and a storage medium, and solves the problems of limited jaw positioning resolution, easy force oscillation, clamping injury, slippage and model prediction distortion when grasping unknown objects in traditional force control. The application controls the humanoid robot upper limb to move the jaw to the pre-grasping position, collects the tactile force, motor current and displacement; detects the first contact to construct the equivalent closed displacement, online identifies the displacement and grasping force polynomial contact model, predicts the grasping force after limiting the local stiffness amplitude; calculates the model confidence and confidence margin by comprehensively considering the residual, sample and contact state, and executes four types of controls including complete adjustment, scaling adjustment, holding and reverse release in layers; resets the model and resamples when the index is abnormal, and maintains the pose and monitors in real time after stable grasping. The application does not need to pre-calibrate the object parameters, restrains the force oscillation through the confidence constraint, considers the grasping flexibility and safety, and is suitable for grasping unknown objects such as soft and slippery objects.

Description

Technical Field

[0001] This invention relates to the field of tactile force control technology for humanoid robot end effectors, specifically to an adaptive grasping control method, system, and storage medium for the upper limb end effector of a humanoid robot. Background Technology

[0002] Humanoid robots are increasingly being used in industrial services, logistics and handling, healthcare and elderly care, home services, and flexible operations. Compared to traditional fixed robotic arms, humanoid robots typically possess multiple degrees of freedom, including shoulder, elbow, and wrist joints, as well as end effectors, enabling them to perform tasks such as approaching, grasping, transporting, and interacting in complex environments. Among these, the end effector, as the actuator that directly contacts the target object, directly impacts the safety and reliability of the robot's operations.

[0003] Existing gripper control methods typically employ fixed closed positions, fixed gripping forces, or traditional integral force control strategies. When gripper positioning resolution is limited, object contact stiffness is unknown, and tactile signals fluctuate, problems such as gripping force oscillation, overpressure injury, unstable gripping, or object slippage can easily occur, reducing the adaptability of humanoid robots to complex grasping objects.

[0004] Therefore, it is necessary to propose an adaptive grasping control method for the upper limb end of a humanoid robot. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive grasping control method, system, and storage medium for the upper limb of a humanoid robot, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An adaptive grasping control method for the upper limb of a humanoid robot includes the following steps:

[0008] S1. Obtain the spatial position of the target object, collect the status information of the humanoid robot's upper limbs (such as shoulder joint, elbow joint, wrist joint, etc.) and end gripper (such as gripper opening and closing position, gripper tactile force and gripper motor current, etc.) in real time, control the humanoid robot's upper limbs to move the end gripper to the pre-grasping posture near the target object, and control the end gripper to gradually close.

[0009] The gripper opening position corresponding to the first contact of the end gripper with the target object is taken as the reference opening position, and the equivalent closing displacement is defined as the difference between the reference opening position and the real-time gripper opening position.

[0010] S2. Based on the equivalent closed displacement and comprehensive tactile force of the gripper, construct the contact response model online, calculate the local contact response slope of the model at the current equivalent closed displacement, and perform amplitude limiting processing on the local contact response slope.

[0011] Combining the current gripping force error of the gripper and the minimum controllable displacement of the gripper, the predicted displacement is calculated. Based on the local contact response slope after amplitude limiting and the predicted displacement, the predicted gripping force for the next cycle is calculated, and the gripping force prediction error is also calculated.

[0012] S3. Construct the model confidence score, which is obtained by weighted summation of the prediction residual confidence score, sample size confidence score, sample distribution confidence score, and contact state stability confidence score. Calculate the confidence margin based on the model confidence score.

[0013] S4. By integrating the gripping force prediction error, model confidence, confidence margin, and safe gripping force threshold, an adaptive decision controller is constructed. The controller performs full output, scaling, holding, or release operations on candidate closed compensation quantities, outputs the actual control compensation quantity, and updates the gripper control input, thereby achieving adaptive and stable gripping of the target object by the end gripper.

[0014] As a further embodiment, the present invention also includes the following steps:

[0015] S5. Collect the real-time gripping force for the next cycle, calculate the predicted residual, and recursively update the moving average residual. When any of the following conditions are met: the moving average residual is greater than the residual threshold, the model confidence is lower than the minimum confidence threshold, the tactile force difference coefficient is greater than the imbalance threshold, or the change in gripper motor current is greater than the current mutation threshold, the contact response model is determined to be in failure, and the model reset operation is performed.

[0016] S6. If the model is not invalid, and the actual grasping force falls within the target grasping force range, and the model confidence is higher than the preset threshold, then the gripper is kept in the closed position to achieve stable grasping. The humanoid robot's upper limbs (such as shoulder joints, elbow joints, and wrist joints) maintain the end-effector spatial pose. During the grasping maintenance process, the tactile force, motor current, and gripper displacement are continuously monitored. If a decrease in grasping force, gripper force imbalance, or sudden change in gripper motor current is detected, then return to step S2 and re-execute adaptive grasping control.

[0017] This invention also proposes an adaptive grasping control system for the upper limb of a humanoid robot, used to execute the aforementioned adaptive grasping control method, including a task and visual input module, an upper limb motion planning and pre-positioning module, an upper limb actuator, an end-effector execution module, a perception and acquisition module, a data fusion and state estimation module, an online identification module for contact response models, a predictive grasping force calculation module, a model confidence evaluation module, an adaptive decision controller, a model update and reset module, and a stable grasping and state output module;

[0018] The sensing and acquisition module includes a tactile detection module, a motor current acquisition module, and a gripper displacement detection module, which collects gripper tactile force, motor current, and gripper opening displacement in real time and transmits them to the data fusion and state estimation module.

[0019] The upper limb motion planning and pre-positioning module receives the target object position and grasping direction output by the task and visual input module, and outputs upper limb motion commands to the upper limb actuator.

[0020] The data fusion and state estimation module performs fusion processing on the data collected by the sensing and acquisition module, and outputs equivalent closed displacement, comprehensive tactile force, differences in tactile force of each claw, and current change to the online identification module of the contact response model.

[0021] The online identification module for the contact response model constructs a contact sample set based on equivalent closed displacement and comprehensive tactile force, fits the contact response model online, and outputs the local contact slope to the prediction gripping force calculation module;

[0022] The predicted gripping force calculation module outputs the predicted gripping force and the gripping force prediction error to the model confidence evaluation module.

[0023] The model confidence evaluation module calculates the model confidence and confidence margin, and transmits them to the adaptive decision controller;

[0024] The adaptive decision controller outputs the actual control compensation amount to the end gripper execution module based on the gripping force prediction error, model confidence, and safe gripping force threshold, so as to realize the end gripper's stable gripping of the target object, and at the same time feeds the control command back to the model update and reset module.

[0025] The model update and reset module completes the recursive update of the prediction residual and the model failure judgment. When the model fails, it outputs a reset command to the contact response model online identification module.

[0026] The stable grasping and status output module receives the stable grasping command output by the adaptive decision controller, continuously monitors the grasping status, and outputs the grasping status, confidence level, and alarm information. When the grasping status is abnormal, it feeds back to the upper limb motion planning and pre-positioning module for re-grasping.

[0027] The present invention also proposes a storage medium storing computer instructions, which, when executed by a processor, perform the adaptive grasping control method as described above.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] 1. This invention establishes an online contact response model between the equivalent closed displacement of the gripper and the comprehensive tactile force, enabling the gripper to adaptively estimate the influence of the control action on the gripping force based on the actual contact process without needing to accurately obtain the material parameters of the target object in advance.

[0030] 2. This invention predicts the gripping force in the next cycle after the candidate closure compensation is applied, thus avoiding the blind closing or releasing of traditional integral force control based solely on the current error, thereby reducing force oscillations of the gripper near the target gripping force in low-positioning-resolution grippers.

[0031] 3. This invention introduces model confidence and confidence margin, and comprehensively considers prediction residuals, sample size, sample distribution range and contact state changes to evaluate the reliability of the contact response model, which can avoid the continuous influence of erroneous models on control decisions;

[0032] 4. This invention expands the control action from the traditional two-level control strategy of execution or cancellation to a multi-level control strategy such as full execution, scaling execution, hold still, reverse release, or model reset, which improves the gripper's adaptability to soft, slippery, and other unknown objects; when the predicted gripping force exceeds the safety threshold, release compensation is executed first, which can reduce the risk of overpressure pinching injury and improve the safety of humanoid robots when gripping soft and fragile objects. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the steps of the control method of the present invention;

[0034] Figure 2 This is a flowchart illustrating the overall control method of the present invention.

[0035] Figure 3 This is a schematic diagram of the module structure of the control system of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Existing gripper control methods typically employ fixed closed positions, fixed gripping forces, or traditional integral force control strategies. Their main drawbacks are as follows:

[0038] First, traditional fixed-position control typically only controls the grippers to close to a preset position, and cannot adaptively adjust the clamping force according to the actual force state of the target object. When the object is hard, it is easy to generate excessive contact force, causing damage to the object or impact to the grippers; when the object is soft or slippery, the clamping force may be insufficient, causing the object to slip or fall off.

[0039] Second, traditional fixed force control or integral force control typically adjusts the gripper control input directly based on the error between the desired gripping force and the actual tactile force. Since grippers usually have limited positioning resolution, the actual displacement is not continuous but constrained by a minimum controllable displacement. When the difference between the target gripping force and the actual gripping force is less than the force change caused by a minimum displacement step, the traditional integral controller may repeatedly close and release near the target force, leading to gripping force oscillations, reduced control accuracy, and unstable gripping state.

[0040] Third, some existing methods predict the gripping force in the next cycle by establishing a positional and force relationship model between the gripper's closing displacement and the tactile force, thereby suppressing oscillations caused by traditional integral control. However, such contact response models are usually obtained by fitting online sampled data, and their accuracy is affected by factors such as the number of samples, the range of sample distribution, changes in contact state, tactile noise, motor current fluctuations, and object slippage. When the contact response model estimation is inaccurate, the predicted gripping force may overestimate or underestimate the actual gripping force, leading to erroneous control decisions, such as continuing to close when it should remain closed, or continuing to apply force when it should release, further causing overpressure, slippage, or oscillations.

[0041] Please see Figures 1-3 This invention proposes an adaptive grasping control method for the upper limb end of a humanoid robot, comprising the following steps:

[0042] S1. Acquire the spatial position of the target object and collect real-time status information of the humanoid robot's upper limbs (such as shoulder joints, elbow joints, and wrist joints) and end gripper (such as gripper opening and closing position, gripper tactile force, and gripper motor current). Based on the spatial position and grasping direction of the target object, control the humanoid robot's upper limbs to move the end gripper to a pre-grasping pose near the target object and ensure that the gripper opening direction is consistent with the target grasping direction. After the gripper approaches the target object, control the end gripper to gradually close.

[0043] In this embodiment, the humanoid robot's upper limb can specifically include the shoulder joint, elbow joint, wrist joint, and connecting limbs between the joints. The state information of the humanoid robot's upper limb can specifically include the position information, displacement information, velocity, acceleration, joint servo torque, and servo motor current, voltage, and other data information of each joint. The state information of the end effector gripper can specifically include the gripper opening and closing position, displacement, velocity, acceleration, limit signal, gripper tactile force, and gripper motor current, voltage, and other data information.

[0044] In this embodiment, during the closing process of the end gripper, the gripper can be controlled to close gradually at a preset closing speed or a preset closing step length, and information such as the tactile force on the left and right sides of the gripper, motor current, and gripper opening position can be collected in real time.

[0045] The gripper opening position corresponding to the first contact of the end gripper with the target object is taken as the reference opening position, and the equivalent closing displacement is defined as the difference between the reference opening position and the real-time gripper opening position.

[0046] Initial contact detection determination and equivalent closed displacement establishment: Define the opening width of the gripper in the kth control cycle as... This refers to the real-time opening position of the gripper; the opening width when the gripper first contacts the target object is... The reference opening position is defined as follows: the gripper is considered to have made initial contact with the target object when any of the following conditions are met:

[0047] T1, ,in, The overall tactile force of the gripper in the k-th control cycle is defined as the actual grasping force of the gripper. The threshold value for tactile force is given by k, which represents the kth cycle.

[0048] T2, ,in, The gripper motor current is the value for the k-th control cycle. This is the no-load current. The current threshold;

[0049] T3 ,in, The gripper's overall tactile force during the (k-1)th control cycle. The time interval between two adjacent cycles. The threshold for the rate of change of tactile force.

[0050] To facilitate the description of the compression process after the gripper contacts the target object, the initial contact position is taken as the reference opening position, and the equivalent closing displacement is defined as follows:

[0051] ;

[0052] in, Let be the equivalent closed displacement in the k-th control cycle. As the reference opening position, The real-time opening position of the gripper during the k-th control cycle;

[0053] As the grippers continue to close... Decrease Increase, therefore It can characterize the relative compression of the gripper on the target object.

[0054] Because the gripper has a limited positioning resolution, its actual displacement is constrained by a minimum controllable displacement; definition The expression relating the real-time opening position of the gripper to the gripper control input, representing the minimum controllable displacement of the gripper, is as follows:

[0055] ;

[0056] in, This represents the minimum controllable displacement of the gripper. This is the current gripper control input.

[0057] S2. Based on the equivalent closed displacement and comprehensive tactile force of the gripper, construct the contact response model online, calculate the local contact response slope of the model at the current equivalent closed displacement, and perform amplitude limiting processing on the local contact response slope.

[0058] Combining the current gripping force error of the gripper and the minimum controllable displacement of the gripper, the predicted displacement is calculated. Based on the local contact response slope after amplitude limiting and the predicted displacement, the predicted gripping force for the next cycle is calculated, and the gripping force prediction error is also calculated.

[0059] Tactile Force Synthesis and Contact State Description: In this embodiment, the gripper specifically has two branch grippers. The tactile forces collected by the tactile sensors of the left and right branch grippers in the kth control cycle are defined as follows: and The combined tactile force of the gripper, also known as the gripping force, is expressed as follows:

[0060] ;

[0061] in, The gripping force of the gripper in the k-th control cycle. For the left tactile force in the kth control cycle, The right tactile force during the kth control cycle;

[0062] To characterize the degree of uneven force distribution between the left and right sides, a tactile force difference coefficient is defined, and its expression is:

[0063] ;

[0064] in, The coefficient of tactile difference. To prevent small constants with a denominator of zero, when When the value exceeds the preset threshold, it indicates that the target object may have eccentric contact, local slippage, or attitude deflection. The system reduces the model confidence and prepares to perform model reset or re-grab.

[0065] Construction of the online identification module for the contact response model: After the gripper makes initial contact with the target object, a contact sample set is constructed based on the equivalent closed displacement and the comprehensive tactile force, and its expression is:

[0066] ;

[0067] in, To access the sample set, Let be the equivalent closed displacement at the j-th sampling time. Let M be the overall tactile force at the j-th sampling time, and M be the current number of samples;

[0068] Based on the aforementioned contact sample set, an online contact response model is established between the equivalent closing displacement of the gripper and the comprehensive tactile force, and its expression is as follows:

[0069] ;

[0070] in, This represents the contact response model obtained through online identification during the k-th control cycle. Let N represent the equivalent closed displacement of order i, where N is the polynomial order. These are the model parameters obtained through online identification;

[0071] The model parameters are solved online using the least squares method, and their expression is as follows:

[0072] ;in, This represents the equivalent closed displacement of order i at the j-th sampling time;

[0073] In other embodiments, the contact response model may also be implemented using a local linear model, a piecewise linear model, a lookup table interpolation model, or a recursive least squares model.

[0074] The expression for the slope of the local contact response of the model at the current equivalent closed displacement in step S2 is:

[0075] ;

[0076] in, This represents the local response coefficient of the change in gripper closing displacement to the change in tactile force under the current contact state, i.e., the local contact response slope. This represents the equivalent closed displacement of the (i-1)th order in the k-th control cycle;

[0077] To avoid excessively large predicted displacements due to excessively small or abnormal local slopes, the expression for limiting the local contact response slope is as follows:

[0078] ;

[0079] in, The slope of the local contact response after limiting. The preset minimum local response slope, The slope of the maximum local response.

[0080] When the object is extremely soft, the slope of the local contact response If the amplitude approaches 0, and no amplitude limiting is applied, the subsequent predicted displacement will be... It can explode in size, with an extremely large candidate closure compensation, easily crushing soft or fragile objects; when the object is extremely hard or the sensor noise changes abruptly, the instantaneous local contact response slope... It will be exceptionally large, and subsequent displacement predictions will be needed. The value will approach 0. If no amplitude limiting is performed, the subsequent controller will not adjust at all, and the gripping force will continue to deviate from the target value, making the object easy to slip and fall off. This invention limits the slope of the local contact response, which can effectively ensure that the predicted step size will not go out of control to an extreme, and ensure that the gripper can adaptively and compliantly grip soft and slippery objects without overpressure.

[0081] The expression for calculating the predicted displacement in step S2 is as follows:

[0082] ;

[0083] ;

[0084] in, Indicates the predicted displacement. Represents a symbolic function. Indicates rounding up. To prevent small constants with a denominator of zero, This represents the minimum controllable displacement of the gripper. This indicates the current gripping force error. The expected grasping force in the k-th control cycle. For the k-th control cycle, the real-time integrated tactile force is used.

[0085] The calculation of predicted displacement takes into account the influence of the minimum controllable displacement step of the gripper on the tactile force, thereby avoiding erroneous predictions under low positioning resolution conditions.

[0086] The formula for calculating the gripping force in the next cycle is:

[0087] ;

[0088] in, Predict the grasping force for the (k+1)th control cycle;

[0089] The formula for calculating the gripping force prediction error is:

[0090] ;

[0091] in, The error in the grasping force prediction during the (k+1)th control cycle is... The expected grasping force for the (k+1)th control cycle.

[0092] S3. Construct the model confidence score, which is obtained by weighted summation of the prediction residual confidence score, sample size confidence score, sample distribution confidence score, and contact state stability confidence score. Calculate the confidence margin based on the model confidence score.

[0093] The process of constructing the model confidence in step S3 is as follows:

[0094] First, considering the predicted residual, after executing the control compensation amount from the previous control cycle, the system obtains the real-time gripper tactile force. Compared to the previous cycle's predicted tactile ability The prediction residuals between the two are:

[0095] ;

[0096] Then, the moving average prediction residual is defined, and its expression is:

[0097] ;

[0098] in, Let L be the moving average prediction residual, and L be the moving window length. Let be the prediction residual at the j-th sampling time;

[0099] Therefore, the confidence level of the predicted residual is constructed, and its expression is:

[0100] The smaller the prediction residual, the more reliable the model.

[0101] in, This is the residual sensitivity coefficient;

[0102] Secondly, considering the sample size, the contact-response model cannot be fully trusted when the sample size is insufficient. Therefore, the sample size confidence score is defined as follows:

[0103] ;

[0104] in, This represents the current number of samples. Minimum number of samples required to enable the contact response model;

[0105] Secondly, considering the sample distribution range, if the samples are concentrated within a very narrow displacement range, the model's prediction reliability near the current displacement or in the extrapolated region is low. The sample distribution reliability is defined as follows:

[0106] ;

[0107] in, This is the preset minimum sample coverage range;

[0108] Finally, considering the stability of the contact state, when there are abrupt changes in the difference in tactile force between the left and right sides, the motor current, or the contact center, it indicates that the target object may deflect, slip, or the contact point may change, and the original contact response model is no longer fully applicable; therefore, a contact state stability confidence score is defined, and its expression is:

[0109] ;

[0110] in, The change in left tactile force. This represents the change in right tactile force. This represents the change in current of the gripper motor. The first weighting coefficient, This is the second weighting coefficient;

[0111] Taking all the above factors into account, the formula for calculating the model confidence score is as follows:

[0112] , ;

[0113] in, For model confidence, To predict the confidence level of the residuals, Confidence level for sample size. To determine the reliability of the sample distribution, For the contact state stability confidence level, Non-negative weights for predicting residual confidence levels, The non-negative weights of the confidence score are the sample size. Assign non-negative weights to the samples based on their reliability. The non-negative weights are used to represent the stability confidence of the contact state.

[0114] This invention calculates the model confidence of the contact response model based on the prediction residual, the number of contact samples, the distribution range of contact samples, and the change in contact state, and takes corresponding gripper actions based on the model confidence, which can avoid incorrect control decisions caused by inaccurate estimation of the contact response model.

[0115] Due to contact response model The predictions are based on online identification using a limited number of samples, and their results may be affected by insufficient sample size, narrow sample distribution range, abrupt changes in contact state, and sensor noise. Therefore, this invention introduces a model confidence level. To characterize the reliability of the current contact response model's prediction results, The closer the value is to 1, the more reliable the current model's prediction. The closer the value is to 0, the greater the potential prediction error of the current model.

[0116] The conditions that each of the above weights must meet are as follows:

[0117] With this setup, the model confidence can be determined by the prediction residuals, the number of contact samples, the distribution range of contact samples, and the stability of the contact state.

[0118] The expression for confidence margin is:

[0119] ;

[0120] in, For confidence margin, This represents the maximum prediction error margin; when At higher levels, Smaller, the controller is close to ordinary predictive control; when At lower levels, If the value increases, the controller automatically enters a conservative control state.

[0121] S4. By integrating the gripping force prediction error, model confidence, confidence margin, and safe gripping force threshold, an adaptive decision controller is constructed. The controller performs full output, scaling, holding, or release operations on candidate closed compensation quantities, outputs the actual control compensation quantity, and updates the gripper control input to achieve adaptive and stable gripping of the target object by the end gripper.

[0122] The actual control compensation quantity output by the adaptive decision controller in step S4 The expression is:

[0123] ;

[0124] in, For high confidence threshold, The low confidence threshold, and satisfying ; For integral gain, This represents the grasping force error for the current cycle. This is the scaling factor. To ensure a safe gripping force threshold, To release directional compensation amount;

[0125] Finally, the updated expression for the gripper control input is:

[0126] ;in, This is the current gripper control input. For the gripper control input in the next cycle, This represents the current actual control compensation amount.

[0127] Traditional integral controllers are susceptible to environmental and model uncertainties, leading to error oscillations and insufficient accuracy. The adaptive decision controller of this invention integrates model confidence. When the model predicts the next clamping force, it assesses the model's confidence level; if the predicted force error... Below the current error And confidence level Greater than the high confidence threshold If the model has high confidence and the prediction results are highly reliable, the action will be executed completely; if the prediction error is high... Below the current error Confidence level Greater than the low confidence threshold And less than the high confidence threshold If the model's confidence level is moderate, scaling is performed; if the prediction error remains unchanged or the model's confidence level is low, no action is taken; if the prediction power exceeds the safety limit, the compensation amount is released in reverse to avoid applying excessive clamping force to the target object.

[0128] The stability and security of the adaptive grasping control method of the present invention will be explained below, assuming that the error between the actual grasping force and the predicted grasping force satisfies the following condition:

[0129] ;

[0130] From the formula ;

[0131] and formula ;

[0132] We can obtain: ;

[0133] When the controller executes the complete control compensation, it satisfies the following condition:

[0134] ;

[0135] Therefore, we can obtain: ;

[0136] Let the Lyapunov function be: ;

[0137] Then we have: The grasping force error decreases monotonically.

[0138] Therefore, if the confidence margin can cover the model prediction error, executing control actions can reduce the grasping force error.

[0139] Meanwhile, if the controller execution conditions meet the following conditions:

[0140] ;

[0141] and: ;

[0142] Then we have: This achieves a safe gripping constraint; therefore, the actual gripping force of this invention will not exceed the safe gripping force threshold; when the predicted gripping force and its confidence margin exceed the safe force threshold, the controller outputs a compensation amount in the release direction, thereby avoiding excessive clamping force on the target object.

[0143] The present invention also includes the following steps:

[0144] S5. Collect the real-time gripping force for the next cycle, calculate the predicted residual, and recursively update the moving average residual. When any of the following conditions are met: the moving average residual is greater than the residual threshold, the model confidence is lower than the minimum confidence threshold, the tactile force difference coefficient is greater than the imbalance threshold, or the change in gripper motor current is greater than the current mutation threshold, the contact response model is determined to be in failure, and the model reset operation is performed.

[0145] As a specific solution, after implementing the control compensation, the system collects the actual grasping force for the next cycle. The predicted residuals are calculated, and their expression is:

[0146] ;in, Predict the residual for the (k+1)th control cycle;

[0147] The sliding residual is updated recursively, and its expression is:

[0148] ;

[0149] in, The moving average residual for the k-th control period is... The moving average residual for the (k+1)th control period is... These are the residual update coefficients;

[0150] The current contact response model is deemed to have failed if any of the following conditions are met:

[0151] ;or ;or ;or ;

[0152] in, To predict the residual threshold, The minimum confidence threshold. The threshold for uneven tactile force between the left and right sides. This is the threshold for sudden changes in motor current.

[0153] When the contact response model fails, the system performs a model reset operation, including clearing or partially forgetting the sample set. Reduce integral gain Reduce the gripper closing step size and re-collect the equivalent closing displacement and tactile force samples of the gripper to re-identify the contact response model.

[0154] S6. If the model is not invalid, and the actual grasping force falls within the target grasping force range, and the model confidence is higher than the preset threshold, then maintain the gripper closed position or use a preset small-amplitude compensation method to maintain the grasping force stability and achieve stable grasping. During the stable holding process, the humanoid robot's upper limbs (such as shoulder joints, elbow joints, and wrist joints) maintain the end-effector spatial pose. During the grasping holding process, the tactile force, motor current, and gripper displacement are continuously monitored. If a decrease in grasping force, gripper force imbalance (i.e., increased difference in tactile force between the left and right sides) or a sudden change in gripper motor current is detected, then return to step S2 and re-execute adaptive grasping control to suppress object slippage or changes in contact state.

[0155] The main principle of this invention is as follows: after the gripper contacts the target object, an online contact response model is established to predict the impact of candidate control actions on the gripping force in the next cycle. Furthermore, model confidence and confidence margin are introduced to evaluate whether the prediction results are reliable. When the model is reliable, control compensation is performed; when the model is not reliable enough, control is scaled or maintained; when there is an overpressure risk, release is performed; and when the model continues to fail, it is reset. This improves the stability, safety, and adaptability of gripping unknown objects, soft objects, and slippery objects.

[0156] The method of this invention can achieve adaptive gripping control of objects with different stiffness, shape and friction characteristics without the need to obtain the object material parameters in advance. It effectively reduces the gripping force oscillation of low positioning resolution grippers in the closed-loop force control process, and improves the stability, safety and environmental adaptability of gripping unknown objects, soft objects and slippery objects.

[0157] This invention also proposes an adaptive grasping control system for the upper limb of a humanoid robot, used to execute the aforementioned adaptive grasping control method, including a task and visual input module, an upper limb motion planning and pre-positioning module, an upper limb actuator, an end-effector execution module, a perception and acquisition module, a data fusion and state estimation module, an online identification module for contact response models, a predictive grasping force calculation module, a model confidence evaluation module, an adaptive decision controller, a model update and reset module, and a stable grasping and state output module;

[0158] The sensing and acquisition module includes a tactile detection module, a motor current acquisition module, and a gripper displacement detection module, which collects gripper tactile force, motor current, and gripper opening displacement in real time and transmits them to the data fusion and state estimation module.

[0159] The upper limb motion planning and pre-positioning module receives the target object position and grasping direction output by the task and visual input module, and outputs upper limb motion commands to the upper limb actuator.

[0160] The data fusion and state estimation module performs fusion processing on the data collected by the sensing and acquisition module, and outputs equivalent closed displacement, comprehensive tactile force, differences in tactile force of each claw, and current change to the online identification module of the contact response model.

[0161] The online identification module for the contact response model constructs a contact sample set based on equivalent closed displacement and comprehensive tactile force, fits the contact response model online, and outputs the local contact slope to the prediction gripping force calculation module;

[0162] The predicted gripping force calculation module outputs the predicted gripping force and the gripping force prediction error to the model confidence evaluation module.

[0163] The model confidence evaluation module calculates the model confidence and confidence margin, and transmits them to the adaptive decision controller;

[0164] The adaptive decision controller outputs the actual control compensation amount to the end gripper execution module based on the gripping force prediction error, model confidence, and safe gripping force threshold, so as to realize the end gripper's stable gripping of the target object, and at the same time feeds the control command back to the model update and reset module.

[0165] The model update and reset module completes the recursive update of the prediction residual and the model failure judgment. When the model fails, it outputs a reset command to the contact response model online identification module.

[0166] The stable grasping and status output module receives the stable grasping command output by the adaptive decision controller, continuously monitors the grasping status, and outputs the grasping status, confidence level, and alarm information. When the grasping status is abnormal, it feeds back to the upper limb motion planning and pre-positioning module for re-grasping.

[0167] The present invention also proposes a storage medium storing computer instructions, which, when executed by a processor, execute the adaptive grasping control method.

[0168] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An adaptive grasping control method for the upper limb end of a humanoid robot, characterized in that, Includes the following steps: S1. Obtain the spatial position of the target object, collect the status information of the humanoid robot's upper limbs and end gripper in real time, control the humanoid robot's upper limbs to move the end gripper to the pre-grasping pose near the target object, and control the end gripper to gradually close. The gripper opening position corresponding to the first contact of the end gripper with the target object is taken as the reference opening position, and the equivalent closing displacement is defined as the difference between the reference opening position and the real-time gripper opening position. S2. Based on the equivalent closed displacement and comprehensive tactile force of the gripper, construct the contact response model online, calculate the local contact response slope of the model at the current equivalent closed displacement, and perform amplitude limiting processing on the local contact response slope. Combining the current gripping force error of the gripper and the minimum controllable displacement of the gripper, the predicted displacement is calculated. Based on the local contact response slope after amplitude limiting and the predicted displacement, the predicted gripping force for the next cycle is calculated, and the gripping force prediction error is also calculated. S3. Construct the model confidence score, which is obtained by weighted summation of the prediction residual confidence score, sample size confidence score, sample distribution confidence score, and contact state stability confidence score. Calculate the confidence margin based on the model confidence score. S4. By integrating the gripping force prediction error, model confidence, confidence margin, and safe gripping force threshold, an adaptive decision controller is constructed. The controller performs full output, scaling, holding, or release operations on candidate closed compensation quantities, outputs the actual control compensation quantity, and updates the gripper control input, thereby achieving adaptive and stable gripping of the target object by the end gripper.

2. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 1, characterized in that, It also includes the following steps: S5. Collect the real-time gripping force for the next cycle, calculate the predicted residual, and recursively update the moving average residual. When any of the following conditions are met: the moving average residual is greater than the residual threshold, the model confidence is lower than the minimum confidence threshold, the tactile force difference coefficient is greater than the imbalance threshold, or the change in gripper motor current is greater than the current mutation threshold, the contact response model is determined to be in failure, and the model reset operation is performed. S6. If the model is not invalid, and the actual grasping force falls within the target grasping force range, and the model confidence is higher than the preset threshold, then maintain the gripper closed position to achieve stable grasping, and the humanoid robot's upper limb maintains the end-effector spatial pose. During the grasping maintenance process, continuously monitor the tactile force, motor current, and gripper displacement. If a decrease in grasping force, gripper force imbalance, or sudden change in gripper motor current is detected, return to step S2 and re-execute adaptive grasping control.

3. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 2, characterized in that, In step S2, the determination that the end gripper makes its first contact with the target object must meet any of the following conditions: T1, ,in, The gripper's overall tactile force during the k-th control cycle is the grasping force of the gripper during the k-th control cycle. The threshold value for tactile force is given by k, which represents the kth cycle. T2, ,in, The gripper motor current is the value for the k-th control cycle. This is the no-load current. The current threshold; T3, ,in, The gripper's overall tactile force during the (k-1)th control cycle. The time interval between two adjacent cycles. The threshold for the rate of change of tactile force.

4. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 3, characterized in that, The specific method for constructing the contact response model in step S2 is as follows: After the gripper makes initial contact with the target object, a contact sample set is constructed based on the equivalent closed displacement and the combined tactile force, and its expression is: ； in, To access the sample set, Let be the equivalent closed displacement at the j-th sampling time. Let M be the overall tactile force at the j-th sampling time, and M be the current number of samples; Based on the aforementioned contact sample set, an online contact response model is established between the equivalent closing displacement of the gripper and the comprehensive tactile force, and its expression is as follows: ； in, This represents the contact response model obtained through online identification during the k-th control cycle. Let N represent the equivalent closed displacement of order i, where N is the polynomial order. These are the model parameters obtained through online identification; The expression for the equivalent closed displacement is: ； in, Let be the equivalent closed displacement during the k-th control cycle. As the reference opening position, The real-time opening position of the gripper during the k-th control cycle; The expression relating the real-time opening position of the gripper to the gripper control input is: ； in, This represents the minimum controllable displacement of the gripper. This is the current gripper control input.

5. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 4, characterized in that, The expression for the slope of the local contact response of the model at the current equivalent closed displacement in step S2 is: ； in, This represents the local response coefficient of the change in gripper closing displacement to the change in tactile force under the current contact state, i.e., the local contact response slope. This represents the equivalent closed displacement of the (i-1)th order in the k-th control cycle; The expression for limiting the slope of the local contact response is: ； in, The slope of the local contact response after limiting. The preset minimum local response slope, The slope of the maximum local response.

6. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 5, characterized in that, The expression for calculating the predicted displacement in step S2 is as follows: ； ； in, Indicates the predicted displacement. Represents a symbolic function. Indicates rounding up. To prevent small constants with a denominator of zero, This represents the minimum controllable displacement of the gripper. This indicates the current gripping force error. The expected grasping force in the k-th control cycle. For the k-th control cycle, the real-time integrated tactile force is used. The formula for calculating the gripping force in the next cycle is: ； in, Predict the grasping force for the (k+1)th control cycle; The formula for calculating the gripping force prediction error is: ； in, The error in the grasping force prediction during the (k+1)th control cycle is... The expected grasping force for the (k+1)th control cycle.

7. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 6, characterized in that, The process of constructing the model confidence in step S3 is as follows: First, considering the predicted residual, after executing the control compensation amount from the previous control cycle, the system obtains the real-time gripper tactile force. Compared to the previous cycle's predicted tactile ability The prediction residuals between the two are: ； Then, the moving average prediction residual is defined, and its expression is: ； in, Let L be the moving average prediction residual, and L be the moving window length. Let be the prediction residual at the j-th sampling time; Therefore, the confidence level of the predicted residual is constructed, and its expression is: ； in, This is the residual sensitivity coefficient; Secondly, considering the sample size, we define the sample size confidence score, which is expressed as follows: ； in, The current number of samples, Minimum number of samples required to enable the contact response model; Next, considering the range of sample distribution, we define the sample distribution reliability, which is expressed as: ； in, This is the preset minimum sample coverage range; Finally, considering the stability of the contact state, the contact state stability confidence score is defined, and its expression is: ； in, The change in left tactile force. This represents the change in right tactile force. This represents the change in current of the gripper motor. The first weighting coefficient, This is the second weighting coefficient; Taking all the above factors into account, the formula for calculating the model confidence score is as follows: ； in, For model confidence, To predict the confidence level of the residuals, Confidence level for sample size. To determine the reliability of the sample distribution, For the contact state stability confidence level, Non-negative weights for predicting residual confidence levels, The non-negative weights of the confidence score are the sample size. Assign non-negative weights to the samples based on their reliability. For the non-negative weights of the contact state stability confidence, the above weights must satisfy the following conditions: ； The expression for confidence margin is: ； in, For confidence margin, This represents the maximum prediction error margin.

8. The adaptive grasping control method for the upper limb of a humanoid robot according to claim 7, characterized in that, The actual control compensation quantity output by the adaptive decision controller in step S4 The expression is: ； in, For high confidence threshold, The low confidence threshold is met. ; For integral gain, This represents the grasping force error for the current cycle. This is the scaling factor. To ensure a safe gripping force threshold, To release directional compensation amount; Finally, the updated expression for the gripper control input is: ;in, This is the current gripper control input. For the gripper control input in the next cycle, This represents the current actual control compensation amount.

9. An adaptive grasping control system for the upper limb of a humanoid robot, used to execute the adaptive grasping control method according to any one of claims 2-8, characterized in that, include: The module includes: task and visual input module, upper limb motion planning and pre-positioning module, upper limb actuator, end gripper actuation module, perception and acquisition module, data fusion and state estimation module, online identification of contact response model module, predictive grasping force calculation module, model confidence evaluation module, adaptive decision controller, model update and reset module, and stable grasping and state output module. The sensing and acquisition module includes a tactile detection module, a motor current acquisition module, and a gripper displacement detection module, which collects gripper tactile force, motor current, and gripper opening displacement in real time and transmits them to the data fusion and state estimation module. The upper limb motion planning and pre-positioning module receives the target object position and grasping direction output by the task and visual input module, and outputs upper limb motion commands to the upper limb actuator. The data fusion and state estimation module performs fusion processing on the data collected by the sensing and acquisition module, and outputs equivalent closed displacement, comprehensive tactile force, differences in tactile force of each claw, and current change to the online identification module of the contact response model. The online identification module for the contact response model constructs a contact sample set based on equivalent closed displacement and comprehensive tactile force, fits the contact response model online, and outputs the local contact slope to the prediction gripping force calculation module; The predicted gripping force calculation module outputs the predicted gripping force and the gripping force prediction error to the model confidence evaluation module. The model confidence evaluation module calculates the model confidence and confidence margin, and transmits them to the adaptive decision controller; The adaptive decision controller outputs the actual control compensation amount to the end gripper execution module based on the gripping force prediction error, model confidence, and safe gripping force threshold, so as to realize the end gripper's stable gripping of the target object, and at the same time feeds the control command back to the model update and reset module. The model update and reset module completes the recursive update of the prediction residual and the model failure judgment. When the model fails, it outputs a reset command to the contact response model online identification module. The stable grasping and status output module receives the stable grasping command output by the adaptive decision controller, continuously monitors the grasping status, and outputs the grasping status, confidence level, and alarm information. When the grasping status is abnormal, it feeds back to the upper limb motion planning and pre-positioning module for re-grasping.

10. A storage medium, characterized in that, It stores computer instructions, which, when executed by a processor, perform the adaptive grasping control method as described in any one of claims 2-8.

Images