A Dexterous Hand Adaptive Grasping Force Control Method Based on Online Contact Increment Recognition
The adaptive grasping force control method based on online contact increment identification solves the problems of slow grasping force establishment and difficulty in coordinating asynchronous multi-finger contact under different contact conditions by dexterous hands. It enables the normal contact force to reach the target value quickly and smoothly, adapting to both soft and rigid contact conditions.
Patent Information
- Application Number
- CN202611576932.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing dexterous hand grasping force control methods are difficult to adapt to different contact conditions. In particular, the grasping force build-up speed is slow under soft and rigid contact conditions, and peak overshoot and grasping force fluctuations are easy to occur. Furthermore, coordination is difficult during multi-finger asynchronous contact.
An adaptive gripping force control method based on online contact increment identification is adopted. By measuring the normal contact force and the active joint state during the contact establishment phase, a smooth normal contact force increment and equivalent indentation increment are constructed. An online contact increment response model is established, and the model parameters and identification reliability are updated in real time. The nonlinear convergence weights are adjusted, monotonic loading constraints are applied, and the results are converted into active joint angle commands.
It improves the dexterity of the hand to adapt to different contact conditions, takes into account the speed of grasping force establishment and the effect of overshoot suppression, adapts to multi-finger asynchronous contact process, and realizes that the normal contact force can quickly and smoothly reach the target value.
Smart Images

Figure CN123060722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic information technology, and more specifically to a method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition. Background Technology
[0002] When a dexterous hand grasps an unknown or semi-unknown object, the material, local shape, contact posture, and transmission characteristics of the finger mechanism of the target object vary. Even if the active joints perform the same angular increment, the changes in normal contact force generated under different contact conditions may be significantly different.
[0003] Existing dexterous hand grasping force control methods typically employ fixed control parameters, preset contact models, or pre-calibrated force-displacement mapping relationships. Under relatively soft contact conditions, the change in normal contact force caused by a unit change in active joint angle is small, which can easily lead to a slow grasping force build-up rate. Under relatively rigid contact conditions, the normal contact force rises rapidly, which can easily cause peak overshoot and grasping force fluctuations. Therefore, the same set of fixed control parameters is difficult to adapt to different contact conditions simultaneously.
[0004] Some existing methods adjust control quantities by estimating environmental stiffness or contact parameters online. However, effective data is limited in the early stages of contact, and the reliability of the identification results is not yet stable. Directly adjusting based on the identification results may still cause overshoot. At the same time, the fingers of a multi-finger dexterous hand usually cannot establish contact simultaneously. There is also a coordination problem between the first contacting finger and the non-contacting finger in maintaining low force and continuing to close, which cannot be well adapted to the asynchronous contact process of multiple fingers. Summary of the Invention
[0005] In view of this, the present invention provides a dexterous hand adaptive grasping force control method based on online contact increment identification, so as to solve the problem that the fixed parameter grasping force control method in the prior art is difficult to adapt to different contact conditions and cannot adapt well to multi-finger asynchronous contact processes.
[0006] An adaptive grasping force control method for a dexterous hand based on online contact incremental recognition includes: S1, the contact establishment phase, uses position control to move each finger, acquires the normal contact force measurement value and active joint state of each finger, determines the contact state based on the normal contact force, keeps the contacted fingers in low-force contact, and keeps the uncontacted fingers closed; after all fingers have established contact, switches to the grasping force adjustment state, smooths the normal contact force measurement value to obtain smooth normal contact force, and constructs the smooth normal contact force increment, smooth normal contact force change rate, and equivalent indentation increment; S2. Based on the smooth normal contact force increment and equivalent indentation increment of the previous control cycle, establish an online contact increment response model to predict the smooth normal contact force increment of the current control cycle, and use the contact force increment memory coefficient and equivalent contact increment gain as parameters to be identified. S3. Calculate the predicted value of the smoothed normal contact force increment for the current control cycle based on the online contact increment response model. Use the difference between the predicted value and the actual value as the model residual. Update the contact force increment memory coefficient and the equivalent contact increment gain using the model residual. Smooth the absolute value of the model residual to obtain the residual smooth value. Update the identification reliability based on the model residual and the effective identification conditions. S4. The relative contact softness is determined based on the updated equivalent contact increment gain, and the relative contact softness and identification reliability are mapped to nonlinear convergence weights. S5. Determine the dynamic prediction time domain based on the residual smoothing value, calculate the current normal contact force error based on the current control target normal contact force and the smoothed normal contact force, and calculate the prediction force error based on the smoothed normal contact force change rate and the dynamic prediction time domain; determine the expected smoothed normal contact force increment for the next control cycle based on the prediction force error and the nonlinear convergence weight. S6. Using the updated online contact incremental response model, the desired smooth normal contact force increment is inversely solved into the equivalent indentation increment without monotonic loading constraints. Then, monotonic loading constraints are applied to the equivalent indentation increment based on the current normal contact force error to obtain the final desired equivalent indentation increment. S7 converts the final expected equivalent indentation increment into the expected active joint angle increment, calculates the active joint command angle for the next control cycle and executes it; after execution, it reacquires the normal contact force measurement value and active joint state of each finger, and performs closed-loop iteration for the next control cycle.
[0007] The dexterous hand adaptive grasping force control method based on online contact incremental identification provided by the present invention has the following beneficial effects: 1. This invention directly utilizes the smoothed normal contact force increment and the equivalent indentation increment to identify the actual contact increment response model online, without needing to obtain the contact stiffness or material type of the target object in advance. This improves the adaptability to different contact conditions and solves the technical problem that existing methods are difficult to apply to unknown objects due to their reliance on prior models.
[0008] 2. This invention updates model parameters and identifies reliability in real time based on model residuals, and creatively uses the identified equivalent contact increment gain to map to relative contact softness, thereby adjusting the nonlinear convergence weights. In soft contact (low gain), a more aggressive force-building action is allowed, while in hard contact (high gain), the inverse indentation naturally decreases, thus balancing the force-building speed and overshoot suppression effect.
[0009] 3. When the smooth normal contact force does not exceed the current control target normal contact force, the present invention applies a monotonic loading constraint to the equivalent indentation increment, so that the finger continues to load or pauses loading without reverse unloading; when the smooth normal contact force exceeds the current control target normal contact force, the monotonic loading constraint is released, allowing the use of negative increment to correct the normal contact force overshoot, thereby reducing repeated adjustments and fluctuations in the gripping force.
[0010] 4. This invention maintains low-force contact for fingers that have already established contact, allows fingers that have not yet established contact to continue closing, and adjusts the grasping force uniformly after all fingers involved in grasping have established contact, thus avoiding the continuous increase in contact force of the fingers that made contact earlier during the waiting process, thereby adapting to the asynchronous contact process of multiple fingers.
[0011] 5. The present invention uses the normal contact force of each finger and the active joint state as control inputs, converts the final expected equivalent indentation increment into an active joint position command, and performs closed-loop iterations in each control cycle through the feedback information after execution, so that the method can be integrated into the existing dexterous hand active joint position control system. Attached Figure Description
[0012] Figure 1 This is a flowchart illustrating a dexterous hand adaptive grasping force control method based on online contact incremental recognition. Figure 2 This is a schematic diagram of the structure of an exemplary three-finger dexterous hand; Figure 3 This is a diagram showing the online identification results of the equivalent contact increment gain and contact force increment memory coefficient of finger 1 under soft contact, medium contact and hard contact conditions in this embodiment of the invention; Figure 4 This is a comparison chart of the actual and predicted values of the smooth normal contact force of finger 1 under soft, medium, and hard contact conditions in an embodiment of the present invention. Figure 5 The simulation curves of the normal contact force response of the method, fixed parameter incremental control, fixed admittance control and PID control of the present invention under soft contact, medium contact and hard contact conditions are shown. Figure 6 This is a simulation curve of three fingers sequentially establishing contact, maintaining low-force contact, and uniformly tracking the target normal contact force in an embodiment of the present invention. Figure 7 The simulation comparison diagram shows the complete method under hard contact conditions, without online model identification, with fixed convergence weights, without monotonic loading constraints, and with a fixed prediction time domain in the embodiments of the present invention. Detailed Implementation
[0013] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.
[0014] Please see Figure 1 The present invention provides a method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition, the method comprising steps S1 to S7: S1, the contact establishment phase, uses position control to move each finger, acquires the normal contact force measurement value and active joint state of each finger, determines the contact state based on the normal contact force, keeps the contacted fingers in low-force contact, and keeps the uncontacted fingers closed; after all fingers have established contact, switches to the grasping force adjustment state, smooths the normal contact force measurement value to obtain smooth normal contact force, and constructs the smooth normal contact force increment, smooth normal contact force change rate, and equivalent indentation increment.
[0015] In this embodiment, please refer to Figure 2 The dexterous hand employs a three-finger structure, with three fingers arranged around the grasping area: finger 1, finger 2, and finger 3. Each finger includes an active joint for receiving position control commands and a coupled phalanx that moves with the active joint. The finger contact detection areas include a proximal phalanx contact detection area 10, a middle phalanx contact detection area 20, and a distal phalanx contact detection area 30. Each contact detection area is equipped with a normal force sensor or tactile sensor to acquire the normal contact force between the corresponding finger and the target object; the active joint status is obtained from feedback signals from a joint encoder, joint angle sensor, or actuator.
[0016] For fingers with passively coupled phalanges, the passively coupled phalanges are not directly controlled but move with the active joint. The controller determines the equivalent indentation change based on the angular change produced by the active joint along the finger's closing direction. It should be noted that... Figure 2 The three-finger dexterous hand shown is only used to illustrate one implementation of this method, and does not impose any restrictions on the number of fingers, link size, or specific coupling form of the dexterous hand.
[0017] Specifically, in this embodiment, during the contact establishment phase, the controller generates active joint position commands cycle by cycle based on the preset active joint approach angular velocity, causing each finger to approach the target object along a preset closing direction. The preset active joint approach angular velocities for fingers 1, 2, and 3 are 34° / s, 29° / s, and 25° / s, respectively. The preset active joint approach angular velocities are set according to the initial posture and closing stroke of each finger, enabling the three fingers to establish contact in different control cycles, thus constituting an asynchronous contact establishment process.
[0018] The aforementioned preset active joint approach angular velocity is the amplitude of the angular velocity along the preset closing direction. In the control calculations of this invention, the finger closing direction is defined as the positive direction, while the joint coordinates used in this embodiment define the finger closing direction as the negative direction. Therefore, as the finger closes, the joint angle gradually decreases from 0°, and the active joint approach angular velocity command takes a negative value. Before the control calculations, this joint angle is converted to an angle with the finger closing direction as the positive direction. In other words, the same finger closing motion is represented as a positive change in the control calculations, but as a negative change in the joint coordinates of this embodiment.
[0019] Within each control cycle, the controller reads the normal contact force measurement value of each finger and applies a connection threshold. and disconnection threshold Determine the contact state of each finger, and .
[0020] When the normal contact force measurement value of the untouched finger is not less than When the finger establishes contact with the target object, it is determined that the finger has made contact; when the measured value of the normal contact force of the contacting finger is lower than the target object, it is determined that the finger has made contact with the target object. When the finger is separated from the target object, it is determined that the finger has separated; when the normal contact force measurement value is located at... and During this period, the contact state of the previous control cycle remains unchanged.
[0021] In this embodiment, the connection threshold The threshold for disconnection is 0.02 N. The value is 0.006 N. By setting different on and off thresholds, frequent switching of the contact state is avoided when the measured value of the normal contact force fluctuates around the contact determination threshold.
[0022] When some fingers have made contact while the others have not, the fingers that have made contact stop performing proximity control and enter a low-force holding state. The preset contact holding force is used as the current control target normal contact force, and the position of the active joint is adjusted according to the error between the current control target normal contact force and the measured value of the normal contact force. The fingers that have not made contact continue to move along the preset closing direction according to the corresponding preset active joint approach angular velocity until contact is established.
[0023] In this embodiment, the preset contact holding force is 0.01 N, which is greater than the disconnection threshold. It is also less than the set target normal contact force, which is used to maintain low-force contact for the already contacted fingers while waiting for the other fingers to establish contact.
[0024] Once all three fingers involved in the grasping action have established contact, each finger ends the contact establishment phase and enters the grasping force adjustment state, switching the current target normal contact force to the set target normal contact force. In this embodiment, the set target normal contact force for each of the three fingers is 2.0 N.
[0025] To reduce the impact of normal contact force measurement noise on the differential calculation of adjacent control cycles, the normal contact force measurement values are smoothed. Specifically, the first... root finger in Smooth normal contact force per control cycle Smooth normal contact force increment Smooth normal contact force change rate They respectively satisfy:
[0026]
[0027]
[0028] in, For the first root finger in Smooth normal contact force for each control cycle; This is the smoothing coefficient for the normal contact force; This is the measured value of the normal contact force; For the first root finger in The smoothed normal contact force variation rate for each control cycle; is the smoothing coefficient for the rate of change of normal contact force, and its value ranges from (0, 1]. To control the cycle.
[0029] Smooth normal contact force change rate Used to describe the changing trend of the current normal contact force; when At that time, it indicates that the normal contact force is increasing; when At that time, it indicates that the normal contact force is decreasing.
[0030] After direction conversion, the equivalent pressure increment satisfy:
[0031] in, For the first root finger in Active joint angle per control cycle; For the first root finger in The active joint angle of each control cycle; the direction of the active joint angle is positive with the direction of finger closure as the positive direction. This indicates an increase in the equivalent pressure. This indicates a reduction in the equivalent pressure.
[0032] In this embodiment, the control cycle Set to 0.001s, normal contact force smoothing coefficient Set to 0.90, smoothing coefficient for normal contact force change rate. Set it to 0.30.
[0033] S2. Based on the smoothed normal contact force increment and equivalent pressure increment of the previous control cycle, an online contact increment response model is established to predict the smoothed normal contact force increment of the current control cycle, and the contact force increment memory coefficient and equivalent contact increment gain are used as parameters to be identified.
[0034] Within each control cycle, the controller uses the normal contact force increment and equivalent pressure increment from the previous control cycle to calculate the predicted normal contact force increment for the current control cycle. Specifically, the online contact increment response model satisfies:
[0035] in, For the first root finger in Predicted value of smoothed normal contact force increment for each control cycle; For the first root finger in The contact force increment memory coefficient for each control cycle is used to characterize the continuation effect of the normal contact force increment of the previous control cycle on the current control cycle. For the first root finger in Smooth normal contact force increment per control cycle; For the first root finger in The equivalent contact increment gain for each control cycle is used to characterize the increase in normal contact force caused by a unit equivalent indentation increment. For the first root finger in The equivalent pressure increment for each control cycle.
[0036] The contact force increment memory coefficient is used to describe the smoothing effect of the normal contact force increment caused by factors such as local deformation of the target object, deformation of the finger structure, contact damping, and actuator dynamics. When the contact force increment memory coefficient is large, the normal contact force change generated in the previous control cycle has a more significant impact on the current control cycle; when the contact force increment memory coefficient is small, the effect of the normal contact force change in the previous control cycle is relatively weak.
[0037] The equivalent contact increment gain is used to describe the incremental conversion relationship between active joint motion and changes in normal contact force. When the equivalent contact increment gain is large, the same equivalent indentation increment can cause a large smooth normal contact force increment; when the equivalent contact increment gain is small, the same equivalent indentation increment causes a relatively small smooth normal contact force increment.
[0038] In this embodiment, the online contact incremental response model does not separately identify the elastic modulus, contact damping, contact point location, and local contact geometry parameters of the target object. Instead, it converts the actual contact response, formed by the combined action of finger mechanism transmission, contact material deformation, and active joint motion, into a contact force incremental memory coefficient and an equivalent contact incremental gain. Therefore, the controller can directly utilize the normal contact force and the active joint state for online updates.
[0039] S3. Calculate the predicted value of the smoothed normal contact force increment for the current control cycle based on the online contact increment response model. Use the difference between the predicted value and the actual value as the model residual. Update the contact force increment memory coefficient and the equivalent contact increment gain using the model residual. Smooth the absolute value of the model residual to obtain the residual smoothing value. Update the identification reliability based on the model residual and the effective identification conditions.
[0040] Among them, the smoothed normal contact force increment of the current control cycle Predicted value of smoothed normal contact force increment The difference between them is used as the model residual. To reduce the impact of the different numerical scales of the smoothing normal contact force increment and the equivalent indentation increment on parameter updates, force scale is utilized. and pressure scale Constructing a normalized update denominator :
[0041]
[0042] in, To prevent updating regular terms with excessively small denominators; When the model residual is small, it means that the predicted value of the online contact incremental response model is close to the smoothed normal contact force increment of the current control cycle; when the model residual is large, it means that there is still a difference between the current model parameters and the actual contact response.
[0043] The contact force increment memory coefficient and the equivalent contact increment gain are corrected based on the model residuals:
[0044]
[0045] in, For the first root finger in Contact force increment memory coefficient for each control cycle; For the first root finger in Equivalent contact increment gain per control cycle; A positive learning rate; and These are the lower and upper boundaries of the contact force increment memory coefficient, respectively. and These are the lower and upper boundaries of the equivalent contact increment gain, respectively. This represents the amplitude limiting function.
[0046] By applying boundary constraints to the updated parameters to be identified, we can avoid the equivalent contact increment gain being too small, which would cause the denominator of the model inverse solution in step S6 to approach zero, and also avoid parameter mutations caused by short-term abnormal data.
[0047] In this embodiment, For 1N, 1deg =1.0×10 -4 Learning rate The initial value is 0.18; the allowable range of the contact force increment memory coefficient is 0 to 0.90, and the allowable range of the equivalent contact increment gain is 0.30 N / deg to 80 N / deg. The initial value of the contact force increment memory coefficient of finger 1 is set to 0.60, and the initial value of the equivalent contact increment gain is set to 50 N / deg; the same initial values of model parameters are used for the three dynamic contact conditions of soft contact, medium contact, and hard contact.
[0048] To evaluate the reliability of the current identification results, the absolute values of the model residuals are smoothed to obtain smoothed residual values:
[0049] in, For the first root finger in The residual smoothing value for each control cycle; For the first root finger in The residual smoothing value for each control cycle; is the residual smoothing coefficient, and ; When the absolute value of the equivalent pressure increment reaches the preset minimum identification step size and the absolute value of the smooth normal contact force increment reaches the preset minimum identification force increment, it is determined that the current control cycle meets the effective identification conditions.
[0050] Specifically, the preset minimum recognition step size is denoted as... , Used to exclude excessively small equivalent indentation increments caused by fluctuations in the active joint angle resolution or numerical values; the preset minimum recognition power increment is denoted as... This is used to eliminate excessively small smoothed normal contact force increments caused by tactile measurement noise. Only when , When the current control cycle is used as a valid identification sample to update the identification reliability, it is determined that the sample meets the valid identification conditions. In this embodiment, 2×10 -5 deg, 2×10 -4 N.
[0051] The identification reliability is limited to the range throughout the entire control cycle. If the aforementioned valid identification conditions are met, the identification reliability is updated as follows:
[0052] in, For the first root finger in The reliability of identification for each control cycle; For the first root finger in The reliability of identification for each control cycle; For reliability update coefficients, and , A positive reliability sensitivity coefficient; The positive force stability term; When the model residual is small relative to the current smoothed normal contact force increment, the exponential term approaches 1, and the identification reliability gradually improves; when the model residual is large, the exponential term decreases, and the identification reliability decreases accordingly.
[0053] If the current control cycle does not meet the effective identification conditions, the identification reliability of the previous control cycle remains unchanged.
[0054] In this embodiment, the residual smoothing coefficient The reliability sensitivity coefficient is 0.80. The value is 1.8, representing the force stability term. The initial value is 0.02N, the initial value of the identification reliability is 0, and the initial value of the residual smoothing value is 0.
[0055] To verify the online identification capability of the present invention for different dynamic contact conditions, three conditions—soft contact, medium contact, and hard contact—were set on the three-finger dexterity hand simulation platform. The contact stiffness and contact damping of each condition are shown in Table 1.
[0056] Table 1
[0057] Under three dynamic contact conditions, the same dexterous hand mechanical structure, target object dimensions, and placement method are used, while maintaining consistent initial states of the active joints, control cycles, target normal contact forces, and controller parameters. During the simulation, the controller does not pre-obtain the contact condition category of the target object, nor does it read the contact stiffness and contact damping of the target object. Instead, it identifies the incremental response relationship corresponding to different contact conditions online based solely on the normal contact force of each finger and the state of the active joints.
[0058] Please see Figure 3 , Figure 3 The online identification results of the equivalent contact increment gain and contact force increment memory coefficient of finger 1 under soft contact, medium contact and hard contact conditions are presented. Figure 3 The three colored curves correspond to soft, medium, and hard contact conditions, respectively. Using the same initial parameter values and controller parameters, the equivalent contact increment gains identified for all three conditions stabilized at approximately 21.8 N / deg, 26.8 N / deg, and 31.5 N / deg, respectively. The contact force increment memory coefficients under soft and medium contact conditions gradually approached zero, while under hard contact conditions, a short-term non-zero change occurred in the initial stage of identification, followed by stabilization at a small positive value. Figure 3 It can be seen that the controller can adjust the model parameters online based on the actual contact response without prior knowledge of the contact condition category; the equivalent contact increment gain forms a distinguishable and stable result as the dynamic contact condition changes.
[0059] Please see Figure 4 , Figure 4 The actual and predicted values of the smooth normal contact force of finger 1 under soft, medium and hard contact conditions are given respectively. Figure 4 The horizontal axis represents simulation time, and the vertical axis represents normal contact force. The solid line represents the actual value of smoothed normal contact force, and the dashed line represents the predicted value of smoothed normal contact force. The enlarged view is used to show the difference between the two curves during the gripping force establishment phase.
[0060] To obtain Figure 4 The smoothed normal contact force prediction value shown is the sum of the smoothed normal contact force increment prediction value of the current control cycle and the smoothed normal contact force of the previous control cycle:
[0061] in, For the first The smoothed normal contact force prediction value for each control cycle. The smoothed normal contact force prediction value is only used to evaluate the prediction effect of the online contact incremental response model and is not involved in model residual calculation, model parameter update, or active joint control command generation.
[0062] Depend on Figure 4 It can be seen that in the initial stage of gripping force establishment, due to the limited number of effective identification samples, there is a short-term difference between the predicted and actual smoothed normal contact force values. As the contact force increment memory coefficient and equivalent contact increment gain are gradually updated, the two curves quickly converge and maintain high consistency in the subsequent gripping force adjustment stage. The root mean square error of prediction under soft, medium, and hard contact conditions is 2.81 × 10⁻⁶. -3 N, 7.27×10 -4 N and 5.64×10 -4 N indicates that the online contact incremental response model can accurately predict changes in normal contact force under different contact conditions.
[0063] S4 determines the relative contact softness based on the updated equivalent contact increment gain, and maps the relative contact softness and identification reliability to nonlinear convergence weights.
[0064] Specifically, the relative contact softness is calculated based on the updated equivalent contact increment gain, and the nonlinear convergence weight is calculated based on the relative contact softness and identification reliability. :
[0065]
[0066] in, For the first root finger in Relative contact softness per control cycle, relative contact softness The material parameters are obtained from the equivalent contact increment gain mapping identified online, rather than from direct sensor measurement. For the first root finger in Nonlinear convergence weights for each control cycle; and These are the lower and upper bounds of the mapping for the equivalent contact increment gain, respectively.
[0067] When the equivalent contact increment gain is small, the smooth normal contact force increment caused by the same equivalent indentation increment is small, and the relative contact softness increases accordingly; when the equivalent contact increment gain is large, the smooth normal contact force increment caused by the same equivalent indentation increment is large, and the relative contact softness decreases accordingly.
[0068] The relative contact softness is obtained by mapping the equivalent contact increment gain obtained online, rather than by material parameters directly measured by tactile sensors, and does not independently determine the direction and amount of change of the active joint angle.
[0069] When the identification reliability is low, the nonlinear convergence weights are kept within a relatively conservative range to reduce excessive active joint adjustment before the model parameters have stabilized. As the identification reliability increases, the nonlinear convergence weights gradually increase, enabling the controller to improve the gripping force establishment speed. Relative contact softness is used to reflect the relative incremental response of the current contact conditions. When the contact is relatively soft, the controller allows for a relatively strong force establishment. When the contact is relatively rigid, the equivalent indentation result naturally decreases, thereby reducing the risk of the normal contact force rising too quickly.
[0070] In this embodiment, the upper boundary of the mapping of the equivalent contact increment gain is... Set to 28.5N / deg, map the lower boundary. Set to 24.0N / deg.
[0071] S5. Determine the dynamic prediction time domain based on the residual smoothing value. Calculate the current normal contact force error based on the current control target normal contact force and the smoothed normal contact force. Combine the smoothed normal contact force change rate and the dynamic prediction time domain to calculate the prediction force error. Determine the expected smoothed normal contact force increment for the next control cycle based on the prediction force error and the nonlinear convergence weight.
[0072] Specifically, the prediction bias weight is calculated based on the residual smoothing value, and the dynamic prediction time domain is determined between the minimum prediction time domain and the maximum prediction time domain. :
[0073]
[0074] in, For the first root finger in Prediction deviation weight for each control cycle; A positive prediction bias scale; and These are the maximum prediction time domain and the minimum prediction time domain, respectively; It is a hyperbolic tangent function; when the residual smoothing value increases, the dynamic prediction time domain increases accordingly, so that the controller can consider the changing trend of the normal contact force earlier when the model uncertainty is high.
[0075] Current normal contact force error satisfy:
[0076] in, The normal contact force of the current control target; Based on the current normal contact force error, the smoothed normal contact force change rate, and the dynamic prediction time domain, the prediction time domain force error is calculated. :
[0077] Then determine the effective predictive power error. :
[0078] when This indicates that the normal contact force, according to its current trend, may exceed the current control target normal contact force, thus effectively predicting the force error. Set to zero, pause the force increment adjustment of the next control cycle; when At that time, the time-domain force error will be predicted. As an effective predictive error Continue to adjust the force increment.
[0079] Then, based on the nonlinear convergence weights and effective prediction error Calculate the convergence ratio :
[0080] in, and These are the minimum convergence ratio and the maximum convergence ratio, respectively. The force error scale is positive.
[0081] When the effective predictive error is large and the identification results are relatively reliable, the convergence ratio increases; when the effective predictive error is small or the identification results are not yet stable, the convergence ratio decreases accordingly.
[0082] Furthermore, the expected normal contact force increment for the next control cycle for: .
[0083] In this embodiment, It is 0.32. It is 0.92. It is 0.20N.
[0084] S6. Using the updated online contact incremental response model, the desired smooth normal contact force increment is inversely solved into the equivalent indentation increment without monotonic loading constraints. Then, monotonic loading constraints are applied to the equivalent indentation increment based on the current normal contact force error to obtain the final desired equivalent indentation increment.
[0085] Specifically, by substituting the expected smoothed normal contact force increment of the next control cycle into the updated online contact increment response model, the equivalent indentation increment without monotonic loading constraints is obtained through inverse solution. :
[0086] The above inverse solution process first subtracts the continuation part of the current smooth normal contact force increment from the expected smooth normal contact force increment, and then uses the equivalent contact increment gain to convert the remaining increment into the equivalent pressure increment.
[0087] The equivalent indentation increment is segmented based on the current normal contact force error to obtain the final expected equivalent indentation increment. : .
[0088] Among them, when When the smooth normal contact force does not exceed the current control target normal contact force, a monotonic loading constraint is applied to the equivalent indentation increment; when When the normal contact force is higher than the current target normal contact force, the monotonic loading constraint is released, allowing the negative incremental correction of the normal contact force overshoot.
[0089] S7 converts the final expected equivalent indentation increment into the expected active joint angle increment, calculates the active joint command angle for the next control cycle and executes it; after execution, it reacquires the normal contact force measurement value and active joint state of each finger, and performs closed-loop iteration for the next control cycle.
[0090] In this embodiment, according to the definition above, with the finger closing direction as the positive direction, the desired increase in the active joint angle is... and the active joint command angle of the next control cycle satisfy:
[0091]
[0092] in, This is the active joint command angle for the current control cycle.
[0093] In this embodiment, the joint coordinates are negatively represented by the finger closing direction. Before sending commands to the active joint position controller, the controller converts the aforementioned active joint command angle to the joint coordinates. Therefore, the converted active joint command angle changes from 0° to the negative direction when the finger closes. This conversion only changes the representation direction of the active joint angle and does not change the control relationship between the equivalent indentation increment and the active joint angle increment.
[0094] The converted active joint command angle for the next control cycle is sent to the active joint position controller for execution. The executed active joint state and normal contact force are then fed back to step S1, and the equivalent push-in increment for the next control cycle is calculated based on the actual change in the executed active joint angle. This forms a closed-loop control process from the acquisition of normal contact force and active joint state, construction of contact increment, online model parameter update, nonlinear convergence weight calculation, dynamic prediction, inverse kinematics of equivalent push-in, to the execution of the active joint command.
[0095] To verify the gripping force control effect of this invention under different dynamic contact conditions, three working conditions were set up: soft contact, medium contact, and hard contact. Please refer to [link / reference]. Figure 5 , Figure 5 The normal contact force response of the method of this invention, fixed parameter incremental control, fixed admittance control, and PID control were compared. In the figure, the horizontal axis represents simulation time, the vertical axis represents normal contact force, and the horizontal dashed line represents the target normal contact force of 2.0 N. Each sub-figure corresponds to soft contact, medium contact, and hard contact conditions, respectively. The rise times of the method of this invention under soft contact, medium contact, and hard contact conditions are 0.0170 s, 0.0144 s, and 0.0142 s, respectively, and the times to reach the steady-state range are 0.0766 s, 0.0620 s, and 0.0576 s, respectively. No significant peak overshoot was observed under any of the three conditions. The rise times of fixed parameter incremental control are 0.0344 s, 0.0248 s, and 0.0194 s, respectively, and the rise times of fixed admittance control are 0.1726 s, 0.1064 s, and 0.0670 s, respectively. The peak overshoot rates of PID control under the three conditions are 0.032%, 4.345%, and 28.065%, respectively. Depend on Figure 5 It can be seen that the method of the present invention can achieve the target normal contact force of 2.0N quickly and stably under different dynamic contact conditions, and can take into account both the gripping force establishment speed and the peak overshoot suppression effect.
[0096] Please see Figure 6 , Figure 6The process of three fingers asynchronously establishing contact and uniformly adjusting the grasping force is presented. In the figure, the horizontal axis represents simulation time, and the vertical axis represents normal contact force. Solid lines of different colors represent the normal contact forces of fingers 1, 2, and 3, respectively, while the horizontal dashed line represents the target normal contact force after unified switching. Fingers 1, 2, and 3 establish contact at 1.104s, 1.192s, and 1.296s, respectively. The first finger to make contact maintains low-force contact during the waiting phase. After finger 3 establishes contact, the current target normal contact force of all three fingers uniformly switches to 2.0N, and all enter the stable range of the target normal contact force at 1.344s. Figure 6 As can be seen, the present invention can avoid the contact force from continuously increasing during the waiting process when the first finger is in contact, and can achieve a continuous switching from the position approach state to the grasping force adjustment state after all fingers have established contact.
[0097] Please see Figure 7 , Figure 7 The normal contact force response under hard contact conditions was further compared using the complete method of this invention, the method without online model identification, the method with fixed convergence weights, the method without monotonic loading constraints, and the method with fixed prediction time domain. In the figure, the horizontal axis represents simulation time, and the vertical axis represents normal contact force. A magnified view is used to show the overshoot, fallback, and convergence differences of each method near the target force. The rise time and the time to enter the ±1% stable range of the complete method of this invention are 0.007s and 0.026s, respectively, with no significant peak overshoot. The response times of the method without online model identification are extended to 0.019s and 0.054s, respectively. The peak normal contact force of the method with fixed convergence weights is 2.026N, with a peak overshoot rate of 1.297%. The cumulative normal contact force fallback of the method without monotonic loading constraints is 0.025N. The peak normal contact force of the method with fixed prediction time domain is 2.326N, with a peak overshoot rate of 16.279%, and a cumulative normal contact force fallback of 0.454N. Figure 7 It is evident that online model identification can improve the speed of grasping force establishment, nonlinear convergence weights and dynamic prediction time domain can reduce peak overshoot and subsequent fluctuations, and monotonic loading constraints can reduce the drop in force value below the target force. All control links work together to ensure that the normal contact force reaches the target value quickly and smoothly.
[0098] In summary, the dexterous hand adaptive grasping force control method based on online contact incremental recognition according to the above embodiments has the following beneficial effects: 1. This invention directly utilizes the smoothed normal contact force increment and the equivalent indentation increment to identify the actual contact increment response model online, without needing to obtain the contact stiffness or material type of the target object in advance. This improves the adaptability to different contact conditions and solves the technical problem that existing methods are difficult to apply to unknown objects due to their reliance on prior models.
[0099] 2. This invention updates model parameters and identifies reliability in real time based on model residuals, and creatively uses the identified equivalent contact increment gain to map to relative contact softness, thereby adjusting the nonlinear convergence weights. In soft contact (low gain), a more aggressive force-building action is allowed, while in hard contact (high gain), the inverse indentation naturally decreases, thus balancing the force-building speed and overshoot suppression effect.
[0100] 3. When the smooth normal contact force does not exceed the current control target normal contact force, the present invention applies a monotonic loading constraint to the equivalent indentation increment, so that the finger continues to load or pauses loading without reverse unloading; when the smooth normal contact force exceeds the current control target normal contact force, the monotonic loading constraint is released, allowing the use of negative increment to correct the normal contact force overshoot, thereby reducing repeated adjustments and fluctuations in the gripping force.
[0101] 4. This invention maintains low-force contact for fingers that have already established contact, allows fingers that have not yet established contact to continue closing, and adjusts the grasping force uniformly after all fingers involved in grasping have established contact, thus avoiding the continuous increase in contact force of the fingers that made contact earlier during the waiting process, thereby adapting to the asynchronous contact process of multiple fingers.
[0102] 5. The present invention uses the normal contact force of each finger and the active joint state as control inputs, converts the final expected equivalent indentation increment into an active joint position command, and performs closed-loop iterations in each control cycle through the feedback information after execution, so that the method can be integrated into the existing dexterous hand active joint position control system.
[0103] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition, characterized in that, include: S1, the contact establishment phase, uses position control to move each finger, acquires the normal contact force measurement value and active joint state of each finger, determines the contact state based on the normal contact force, keeps the contacted fingers in low-force contact, and keeps the uncontacted fingers closed; after all fingers have established contact, switches to the grasping force adjustment state, smooths the normal contact force measurement value to obtain smooth normal contact force, and constructs the smooth normal contact force increment, smooth normal contact force change rate, and equivalent indentation increment; S2. Based on the smooth normal contact force increment and equivalent indentation increment of the previous control cycle, establish an online contact increment response model to predict the smooth normal contact force increment of the current control cycle, and use the contact force increment memory coefficient and equivalent contact increment gain as parameters to be identified. S3. Calculate the predicted value of the smoothed normal contact force increment for the current control cycle based on the online contact increment response model. Use the difference between the predicted value and the actual value as the model residual. Update the contact force increment memory coefficient and the equivalent contact increment gain using the model residual. Smooth the absolute value of the model residual to obtain the residual smooth value. Update the identification reliability based on the model residual and the effective identification conditions. S4. The relative contact softness is determined based on the updated equivalent contact increment gain, and the relative contact softness and identification reliability are mapped to nonlinear convergence weights. S5. Determine the dynamic prediction time domain based on the residual smoothing value, calculate the current normal contact force error based on the current control target normal contact force and the smoothed normal contact force, and calculate the prediction force error based on the smoothed normal contact force change rate and the dynamic prediction time domain; determine the expected smoothed normal contact force increment for the next control cycle based on the prediction force error and the nonlinear convergence weight. S6. Using the updated online contact incremental response model, the desired smooth normal contact force increment is inversely solved into the equivalent indentation increment without monotonic loading constraints. Then, monotonic loading constraints are applied to the equivalent indentation increment based on the current normal contact force error to obtain the final desired equivalent indentation increment. S7, convert the final expected equivalent push-in increment into the expected active joint angle increment, calculate the active joint command angle for the next control cycle and execute it; After execution, the normal contact force measurement value of each finger and the active joint state are reacquired, and the closed-loop iteration of the next control cycle is performed.
2. The method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition according to claim 1, characterized in that, In step S1, the connection threshold is used. and disconnection threshold Determine the contact state of each finger, and When the measured value of the normal contact force without touching the finger is not less than When the finger establishes contact with the target object, it is determined that the finger has made contact; when the measured value of the normal contact force of the contacting finger is lower than the target object, it is determined that the finger has made contact with the target object. When the finger is separated from the target object, it is determined that the finger has separated; when the normal contact force measurement value is located at... and During this period, the contact state of the previous control cycle remains unchanged.
3. The method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition according to claim 2, characterized in that, In step S1, the first The root finger in the first Smooth normal contact force per control cycle Smooth normal contact force increment Smooth normal contact force change rate Equivalent pressure increment They respectively satisfy: in, For the first The root finger in the first Smooth normal contact force for each control cycle; This is the smoothing coefficient for the normal contact force; This is the measured value of the normal contact force; For the first The root finger in the first The smoothed normal contact force variation rate for each control cycle; This is the smoothing coefficient for the rate of change of normal contact force; To control the cycle; For the first The root finger in the first Active joint angle per control cycle; For the first The root finger in the first Active joint angle per control cycle.
4. The method for adaptive grasping force control of a dexterous hand based on online contact increment recognition according to claim 3, characterized in that, In step S2, the online contact incremental response model satisfies: in, For the first The root finger in the first Predicted value of smoothed normal contact force increment for each control cycle; For the first The root finger in the first Contact force increment memory coefficient for each control cycle; For the first The root finger in the first Smooth normal contact force increment per control cycle; For the first The root finger in the first Equivalent contact increment gain per control cycle; For the first The root finger in the first The equivalent pressure increment for each control cycle.
5. The method for adaptive grasping force control of a dexterous hand based on online contact increment recognition according to claim 4, characterized in that, In step S3, the smoothed normal contact force increment of the current control cycle is... Predicted value of smoothed normal contact force increment The difference between them is used as the model residual. Utilizing force scale and pressure scale Constructing a normalized update denominator : in, To prevent updating regular terms with excessively small denominators; The contact force increment memory coefficient and the equivalent contact increment gain are corrected based on the model residuals: in, For the first The root finger in the first Contact force increment memory coefficient for each control cycle; For the first The root finger in the first Equivalent contact increment gain per control cycle; A positive learning rate; and These are the lower and upper boundaries of the contact force increment memory coefficient, respectively. and These are the lower and upper boundaries of the equivalent contact increment gain, respectively. This represents the amplitude limiting function.
6. The method for adaptive grasping force control of a dexterous hand based on online contact increment recognition according to claim 5, characterized in that, In step S3, the absolute values of the model residuals are smoothed to obtain smoothed residual values: in, For the first The root finger in the first The residual smoothing value for each control cycle; For the first The root finger in the first The residual smoothing value for each control cycle; This is the residual smoothing coefficient; When the absolute value of the equivalent press-in increment reaches the preset minimum identification step size and the absolute value of the smooth normal contact force increment reaches the preset minimum identification force increment, it is determined that the current control cycle meets the effective identification conditions. The identification reliability is limited to the range throughout the entire control cycle. If the aforementioned valid identification conditions are met, the identification reliability is updated as follows: in, For the first The root finger in the first The reliability of identification for each control cycle; For the first The root finger in the first The reliability of identification for each control cycle; For reliability update coefficients, and , A positive reliability sensitivity coefficient; The positive force stability term; If the current control cycle does not meet the effective identification conditions, the identification reliability of the previous control cycle remains unchanged.
7. The method for adaptive grasping force control of a dexterous hand based on online contact incremental identification according to claim 6, characterized in that, In step S4, the relative contact softness is calculated based on the updated equivalent contact increment gain, and the nonlinear convergence weight is calculated based on the relative contact softness and the identification reliability. : in, For the first The root finger in the first Relative contact softness per control cycle; For the first The root finger in the first Nonlinear convergence weights for each control cycle; and These are the lower and upper bounds of the mapping for the equivalent contact increment gain, respectively.
8. The method for adaptive grasping force control of a dexterous hand based on online contact incremental recognition according to claim 7, characterized in that, In step S5, the prediction bias weight is calculated based on the residual smoothing value, and the dynamic prediction time domain is determined between the minimum prediction time domain and the maximum prediction time domain. : in, For the first The root finger in the first Prediction deviation weight for each control cycle; A positive prediction bias scale; and These are the maximum prediction time domain and the minimum prediction time domain, respectively; It is the hyperbolic tangent function; Current normal contact force error satisfy: in, The normal contact force of the current control target; Based on the current normal contact force error, the smoothed normal contact force change rate, and the dynamic prediction time domain, the prediction time domain force error is calculated. : Then determine the effective predictive power error. : Then, based on the nonlinear convergence weights and effective prediction error Calculate the convergence ratio : in, and These are the minimum convergence ratio and the maximum convergence ratio, respectively. A positive force error scale; Furthermore, the expected normal contact force increment for the next control cycle for: 。 9. The method for adaptive grasping force control of a dexterous hand based on online contact incremental identification according to claim 8, characterized in that, In step S6, the expected smoothed normal contact force increment of the next control cycle is substituted into the updated online contact incremental response model, and the equivalent indentation increment without monotonic loading constraints is obtained by inverse solution. : The equivalent indentation increment is segmented based on the current normal contact force error to obtain the final expected equivalent indentation increment. : 。 10. The method for adaptive grasping force control of a dexterous hand based on online contact incremental identification according to claim 9, characterized in that, In step S7, the desired increase in the active joint angle is... and the active joint command angle of the next control cycle satisfy: in, This is the active joint command angle for the current control cycle.
