Robot remote interaction method and system based on tactile feedback

Abstract

The invention relates to the technical field of robot remote interaction, and discloses a robot remote interaction method and system based on tactile feedback, and the method comprises the steps: obtaining force data, position data and material friction data of interaction between a remote robot and an environment; performing time domain registration and differential processing to obtain an initial contact signal, and extracting a force component to generate a regular contact signal after correction; if the super-stiffness threshold value is the dynamic contact type, constructing a multi-dimensional input vector in combination with the contact dynamic parameter and the friction data, and substituting the multi-dimensional input vector into a pre-training model to generate a predicted contact force; adjusting the friction characteristic weight to obtain a correction feedback signal when the deviation threshold value is exceeded; the high-frequency texture and low-frequency contact force features are decomposed, a composite tactile driving signal is generated, and an execution mechanism is driven to achieve high-fidelity force tactile output. According to the method, high-fidelity haptic feedback of remote interaction of the robot can be achieved, and the strict requirements for remote operation accuracy and safety are met.

Description

Technical Field

[0001] This invention relates to the field of robot remote interaction technology, and in particular to a robot remote interaction method and system based on tactile feedback. Background Technology

[0002] Currently, in the field of robot remote interaction technology, with the continuous expansion of remote operation scenarios and the increasing demand for precision operation, robot remote interaction based on tactile feedback serves as a core technology support. The authenticity and accuracy of its feedback are directly related to the safety and precision of the operation. The efficient analysis and processing capabilities of intelligent chips for multi-source data can support the achievement of this goal.

[0003] In existing technologies, tactile feedback methods for remote robot interaction mainly rely on fixed parameter mapping or simple force transmission. For example, they may directly amplify and output remote force data, or use a pre-set friction prediction model to ignore dynamic changes in contact characteristics, or lack effective processing of signal noise and latency. However, this approach is clearly insufficient in complex interaction scenarios. Fixed parameters cannot adapt to dynamic changes in different materials and contact angles, leading to distorted feedback force and directional deviations. Simple force transmission struggles to filter sensor noise and transmission delays, affecting operational judgment. Furthermore, the lack of real-time tracking and model iteration of contact stiffness and friction characteristics, especially in precision medicine and hazardous environments, can easily lead to excessive force or operational failure, failing to meet the demands of high-precision interaction.

[0004] In summary, existing technologies are insufficient to achieve high-fidelity tactile feedback for remote robot interaction, and cannot meet the stringent requirements for accuracy and safety in remote operations. Summary of the Invention

[0005] This invention provides a method and system for remote robot interaction based on tactile feedback, so as to achieve high-fidelity tactile feedback for remote robot interaction and meet the stringent requirements for accuracy and safety in remote operations.

[0006] In a first aspect, to address the aforementioned technical problems, the present invention provides a method for remote robot interaction based on haptic feedback, comprising:

[0007] Acquire force data, position data, and material friction data of the interaction between the remote robot and its environment; The force data and the position data are time-domain registered and differentiated to obtain the initial contact signal; A state vector is constructed based on the initial contact signal, and a predicted state vector is obtained by substituting it into a preset state transition matrix. Real-time force observations are collected and the predicted state vector is corrected based on the observations. The force components of the corrected state vector are extracted to obtain a regular contact signal. Extract the stiffness change value from the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, analyze the amplitude and frequency of the regular contact signal and compare it with the pre-acquired contact type template to identify the dynamic contact type. Collect the dynamic parameters of the current interaction, combine the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitute the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force; The deviation between the predicted contact force and the actual contact force is calculated. If the deviation exceeds the preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements, and then a correction feedback signal is obtained. The correction feedback signal is decomposed into features. Based on the decomposed high-frequency texture features and low-frequency contact force features, a composite tactile driving signal is generated and executed to obtain a force tactile output consistent with the actual interaction.

[0008] Secondly, the present invention provides a robot remote interaction system based on haptic feedback, comprising: The data acquisition module is used to acquire force data, position data, and material friction data of the interaction between the remote robot and the environment; The contact signal module is used to perform time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal. The signal processing module is used to construct a state vector based on the initial contact signal, and substitute it into a preset state transition matrix to obtain a predicted state vector. It also collects real-time force observations and corrects the predicted state vector based on the observations, extracts the force components of the corrected state vector, and obtains a regular contact signal. The contact judgment module is used to extract the stiffness change value in the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, the amplitude and frequency of the regular contact signal are analyzed and compared with the pre-acquired contact type template to identify the dynamic contact type. The contact force prediction module collects the dynamic contact parameters of the current interaction, combines the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitutes the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force. The signal correction module calculates the deviation between the predicted contact force and the actual contact force. If the deviation exceeds a preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements, and then a correction feedback signal is obtained. The tactile output module performs feature decomposition on the correction feedback signal, generates a composite tactile drive signal based on the decomposed high-frequency texture features and low-frequency contact force features, and executes it to obtain a force tactile output consistent with the actual interaction.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains force data, position data and material friction data of remote robot interaction with environment, performs time domain registration and differential processing on force data and position data, and obtains regular contact signal by combining state transition matrix and real-time force observation value correction. It breaks through the limitation of traditional simple force value transmission that cannot filter noise and delay, explores the temporal characteristics and force-position correlation law of contact process, eliminates sensor noise and transmission delay interference, provides high-precision basic data support for tactile feedback, effectively improves the regularity and reliability of initial contact signal, and solves the problem of operation judgment deviation caused by signal distortion.

[0010] (2) This invention extracts the stiffness change value of the regular contact signal, analyzes the signal amplitude and frequency to identify the dynamic contact type when the threshold is exceeded, and constructs a multi-dimensional model input vector by combining the contact dynamic parameters and material friction data. It is then substituted into the pre-trained friction prediction model to generate the predicted contact force. This invention breaks through the limitation that the traditional fixed parameter mapping cannot adapt to the dynamic contact characteristics, accurately captures the differences in friction characteristics of different contact types and materials, provides multi-dimensional basis for force feedback optimization, significantly improves the accuracy of contact force prediction in complex interaction scenarios, and makes up for the shortcomings of existing technologies that are difficult to adapt to changes in material and contact state.

[0011] (3) This invention optimizes the predicted contact force through the friction prediction model iteratively. When the deviation threshold is exceeded, the model friction characteristic weight is adjusted to generate a correction feedback signal. The high-frequency texture and low-frequency contact force features are decomposed to generate a composite tactile driving signal. This improves the problem of traditional lack of real-time iteration and feature decomposition, provides high-fidelity force tactile feedback for operators, solves the problem of feedback distortion and directional deviation leading to excessive force or operation failure, takes into account the realism of interaction and the safety of operation, and meets the stringent requirements of precision operation for the accuracy of remote interaction. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a robot remote interaction method based on tactile feedback provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a robot remote interaction system based on tactile feedback provided in the second embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for remote robot interaction based on haptic feedback, comprising the following steps: S101, acquires force data, position data, and material friction data of the interaction between the remote robot and the environment; S102, perform time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal; S103, construct a state vector based on the initial contact signal, and substitute it into a preset state transition matrix to obtain a predicted state vector. Collect real-time force observations and correct the predicted state vector based on the observations. Extract the force components of the corrected state vector to obtain a regular contact signal. S104, extract the stiffness change value in the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, analyze the amplitude and frequency of the regular contact signal and compare it with the pre-acquired contact type template to identify the dynamic contact type. S105, Collect the dynamic contact parameters of the current interaction, combine the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitute the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force; S106, calculate the deviation between the predicted contact force and the actual contact force. If the deviation exceeds the preset deviation judgment threshold, adjust the friction characteristic weight of the friction prediction model, regenerate the predicted contact force and perform deviation verification again until the deviation meets the requirements, and obtain the correction feedback signal. S107, the correction feedback signal is decomposed into features, and a composite tactile driving signal is generated and executed based on the decomposed high-frequency texture features and low-frequency contact force features to obtain a force tactile output consistent with the actual interaction.

[0015] In step S101, acquiring force data, position data, and material friction data of the remote robot's interaction with the environment includes: Force data of the interaction between the remote robot and its environment is collected using force sensors; The position data of the remote robot is collected using a position encoder; Retrieve pre-stored material friction data, which includes basic friction factors and surface texture information for different materials.

[0016] It should be noted that, firstly, when collecting force data on the interaction between the remote robot and its environment using a force sensor, the ATIMini45 six-dimensional force sensor is selected. This sensor has a range of 0-45 N (force) and 0-1.8 N·m (torque), a sampling frequency of 1000 Hz, and a measurement accuracy of ±0.01 N. It is installed between the robot's end effector and the tool. During data acquisition, data is transmitted via an EtherCAT bus, and the timestamp of each sampling point is recorded synchronously (accuracy 1 μs) to ensure timing alignment with subsequent position data. After acquisition, the force data is normalized, mapping the force values ​​to the [-1, 1] interval (torque is mapped to the [-0.04, 0.04] interval), eliminating numerical differences caused by different ranges. For example, when the robot grips a plastic part, the X-axis force value collected by the sensor is 5.2 N, which, after normalization, becomes 0.116, clearly reflecting the magnitude of the force during the interaction.

[0017] In this embodiment, when acquiring the position data of the remote robot, a Heidenhain EQN1325 incremental photoelectric encoder with a resolution of 1024 lines / revolution and a sampling frequency of 500Hz is used, installed on the output shaft of each joint motor of the robot. The acquired data is the joint rotation angle pulse signal, which is converted into the actual angle value by a counter, and then combined with the forward solving of the robot's DH parameters to obtain the three-dimensional position coordinates (Cartesian coordinate system) of the end effector. This is then normalized to the [0,1] interval. The volume normalization method adopts a minimum-maximum linear mapping based on the workspace range, and the mapping relationship is set according to the robot's workspace range. For example, the workspace range of the X-axis is set to [-0.1m, 0.9m] (total travel 1.0m). When the robot's end effector moves from (0.2m, 0.3m, 0.4m) to (0.25m, 0.3m, 0.4m), for the X-axis coordinate: the normalized value of the initial position 0.2m is (0.2 - (-0.1)) / 1.0 = 0.3; the normalized value of the position after moving is (0.25 - (-0.1)) / 1.0 = 0.35.

[0018] It should be noted that when retrieving pre-stored material friction data, the material friction data is stored on the SSD hard drive of the robot's local industrial computer, and is classified and stored in JSON format, divided by material type such as metal, plastic, rubber, etc.

[0019] In this embodiment, data construction is achieved through experimental acquisition. Friction experiments are conducted on samples of different materials (three replicates for each material) using a tensile testing machine to measure the sliding friction force under different normal forces (0-50N), and the basic friction factor (the ratio of friction force to normal force) is calculated. Simultaneously, a laser roughness meter (measuring range 0.01-10μm, accuracy 0.001μm) is used to collect surface texture information (average roughness Ra, maximum profile height Rz) of the samples. This data is then integrated to form complete friction data for each material. For example, friction data for plastic material is retrieved, with a basic friction factor of 0.32 and an average surface roughness Ra of 0.8μm, providing basic parameters for subsequent friction prediction model construction.

[0020] In step S102, the step of performing time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal includes: Align the force data and the position data according to the timestamp to generate a synchronization data frame; The synchronous data frame is differentiated to calculate the gradient of force change; If the force change gradient exceeds a preset gradient determination threshold, the corresponding synchronous data frame is locked as instantaneous contact data. The instantaneous contact data is extracted and combined with the position data to generate an initial contact signal.

[0021] It should be noted that, firstly, when generating synchronized data frames by aligning force and position data according to timestamps, the input force and position data both originate from step S101 and have already undergone normalization, so no further dimensional conversion is required. Alignment employs linear interpolation, using the force data timestamp as a reference to interpolate and fill in the position data, ensuring that each force data sampling point has corresponding position data. The time interval between synchronized data frames is uniformly 1ms, and each data frame contains the force value, three-dimensional position coordinates, and a unified timestamp.

[0022] For example, the timestamp of a certain sampling point of force data is 1620000001.001s, and the timestamp of the corresponding location data is 1620000001.0012s. The location coordinates corresponding to the force data are calculated by linear interpolation, and a complete synchronization data frame is generated.

[0023] If the force change gradient exceeds a preset gradient judgment threshold, the corresponding synchronous data frame is locked as instantaneous contact data. In this embodiment, the gradient judgment threshold is set based on the normalized force gradient statistics of interactions with different materials over the past year, maintaining consistency with the dimensions of the input data. The minimum normalized force change gradient of all valid contact scenarios is statistically analyzed, with a base threshold set to 1.1. When interacting with soft materials (such as rubber and fabric), where the gradient change is gradual, the threshold can be lowered to 0.6; when interacting with hard materials (such as metal and glass), where the gradient change is abrupt, the threshold can be raised to 1.5. This threshold can accurately distinguish between real contact and force fluctuations caused by noise. During comparison, the currently calculated force change gradient is compared with this dynamically generated threshold. If it exceeds the threshold, it is marked as contact data. This mechanism ensures that the system maintains optimal contact detection accuracy under different operating conditions.

[0024] Finally, when generating the initial contact signal by extracting the instantaneous contact data and combining it with the position data, a direct feature extraction and vector combination method is used. Specifically, the normalized force value and force change gradient in the synchronous data frame locked as instantaneous contact data are directly read. At the same time, the real-time position coordinates corresponding to this frame are obtained, and the difference (i.e., relative displacement) between them and the contact starting point reference position locked in the previous step in the force direction is calculated to obtain the penetration depth estimate. Finally, the normalized force value, penetration depth estimate, and force change gradient are concatenated in a preset order to construct the initial contact signal in the form of a three-dimensional vector.

[0025] For example, in a locked frame of instantaneous contact data, the normalized force value is 0.062, and the synchronously calculated force change gradient is 11.0. The real-time normalized position corresponding to this frame is 0.303, while the reference position of the contact initiation point is 0.300. Therefore, the estimated penetration depth is 0.303 - 0.300 = 0.003. The final generated initial contact signal is [0.062, 0.003, 11.0]. This signal directly characterizes the mechanical strength, spatial depth, and rate of change at the instant of contact. The data is pure and computationally efficient, providing accurate input features for subsequent state prediction.

[0026] In step S103, the process of constructing a state vector based on the initial contact signal, substituting it into a preset state transition matrix to obtain a predicted state vector, acquiring real-time force observations and correcting the predicted state vector based on the observations, extracting the force components of the corrected state vector, and obtaining a regularized contact signal includes: A state vector is constructed based on the initial contact signal, and the predicted state vector is obtained by substituting it into a preset state transition matrix. Collect real-time force observations, extract the timestamps of the observations and the extrapolation timestamps of the predicted state vectors, calculate the time difference, and generate a delay compensation factor based on the time difference; The observed values ​​are corrected according to the delay compensation factor to obtain aligned observation data; Based on a preset noise suppression coefficient, the aligned observation data and the predicted state vector are weighted and fused, and sensor noise is filtered out through signal smoothing to obtain a corrected state vector. Force component values ​​are extracted from the corrected state vector to obtain a regular contact signal.

[0027] It should be noted that, firstly, when constructing a state vector based on the initial contact signal and substituting it into a preset state transition matrix to derive the predicted state vector, the construction is performed directly using data splicing. Specifically, the three data points contained in the initial contact signal generated in step S102—the normalized force value, the estimated penetration depth, and the force change gradient—are directly spliced ​​together in a preset order to form a three-dimensional state vector. The state vector is set to 3 dimensions, and all components within the vector use the normalization standard from the previous step (i.e., mapped to the [0,1] interval) to ensure dimensional uniformity.

[0028] It is worth noting that the state transition matrix is ​​constructed based on a spring-damping physical model. The matrix elements are calibrated using force-displacement data from interactions with different materials over the past year. For example, the basic matrix parameters are [[0.98,0.01,0.01],[0.01,0.99,0.005],[0.005,0.01,0.97]]. When interacting with soft materials, the damping coefficient can be adjusted to decrease the diagonal elements of the matrix by 0.02, and when interacting with hard materials, it can be increased by 0.01. During the derivation, the state vector is multiplied by the transition matrix to obtain the predicted state vector for the next moment, reflecting the evolution of the contact state without external disturbance.

[0029] For example, the normalized force value in the initial contact signal is 0.3 (corresponding to a physical value of 13.5N), the normalized penetration depth is 0.2, and the normalized force change gradient transmitted in step S102 is 0.08. The state vector [0.3, 0.2, 0.08] is constructed. After substituting into the transition matrix, the predicted state vector is [0.296, 0.201, 0.081], indicating a slight decrease in the force value.

[0030] Next, when generating the delay compensation factor, real-time force observations are acquired using an ATI Mini45 six-dimensional force sensor. After extracting the timestamps of the observed and projected values, the delay compensation factor is calculated by multiplying the absolute time difference by a predetermined base compensation coefficient and then adding 1. The base compensation coefficient is set to 0.001 by default. This coefficient is based on statistical analysis of the effect of transmission delay on signal amplitude attenuation: regression analysis of test data revealed that in a remote interactive environment, for every 1ms increase in transmission delay, the effective amplitude of the signal drifts by an average of approximately 0.1%. Therefore, this coefficient is set to linearly compensate for information lag loss caused by delay. To ensure that the larger the delay, the stronger the compensation, and considering that the absolute time difference is always non-negative, the range of the compensation factor is limited to the interval [1.0, 1.05] (i.e., only an upper limit is set to prevent overcompensation). For example, if the observed timestamp lags behind the projected timestamp by 6ms, the calculated absolute time difference is 6, and the delay compensation factor is 1 + 6 × 0.001 = 1.006, which is within a reasonable range.

[0031] In this embodiment, when obtaining aligned observation data by correcting the observation values ​​according to the delay compensation factor, the collected real-time force observation values ​​are first normalized according to the sensor range (45N) described in step S101 to obtain normalized observation values. Subsequently, the normalized observation values ​​are multiplied by the delay compensation factor to achieve time-domain alignment and amplitude calibration, ensuring that the timestamps and amplitude trends of the observation data are consistent with those of the predicted state vector. Before correction, outliers are removed from the observation values ​​(values ​​exceeding the mean plus or minus 3 times the standard deviation are removed), and then compensation calculation is performed. The aligned data is normalized to the [0,1] interval, maintaining consistency with the state vector dimension.

[0032] For example, the real-time force observation is normalized to 0.32, the delay compensation factor is 1.006, and the corrected aligned observation data is 0.32×1.006≈0.322 (normalized to 0.322), which matches the timestamp and amplitude trend of the predicted state vector.

[0033] It should be noted that when generating the corrected state vector, the noise suppression coefficient Ks is set based on the ratio of the sensor's static noise variance to the dynamic interaction process variance. Based on the Kalman filter gain principle, and considering the proportional relationship between the measurement noise covariance and the state estimation covariance, a base value for this coefficient is set. Experimentally, the base value is set to 0.3. When significant environmental interference (such as strong electromagnetic interference) leads to a substantial increase in measurement noise, the coefficient is increased to 0.4 to rely more heavily on model predictions to suppress observation noise. Conversely, when the environment is stable and interference is low, it is decreased to 0.2 to more sensitively respond to changes in real-time observation data. The residual is the difference between the aligned observation data and the corresponding components of the predicted state vector. During weighted fusion, the weight of the predicted state vector is Ks, and the weight of the aligned observation data is 1-Ks. The fusion yields the initial correction value.

[0034] It is worth noting that the signal smoothing uses a moving average filter with a window size of 5 data points to filter out high-frequency noise. The corrected state vector is the result of smoothing the fused values, with each component still remaining in the [0,1] range.

[0035] For example, the force component of the aligned observation data is 0.322, the force component of the predicted state vector is 0.296, the residual is 0.026, the noise suppression coefficient is 0.3, the weighted fusion force component is 0.296×0.3+0.322×0.7≈0.314, the force component after moving average filtering is 0.31, and the final corrected force component of the state vector is 0.31. The penetration depth and force change gradient are calculated according to the same logic.

[0036] Finally, when extracting the force component values ​​from the corrected state vector to obtain the regularized contact signal, the first element (force component) of the corrected state vector is directly extracted, and denormalized according to the force sensor range (i.e. [0,45N]) set in step S101 to obtain the actual physical force value. Then, the values ​​are arranged in the order of timestamps to form a continuous regularized contact signal.

[0037] For example, the normalized value of the force component of the corrected state vector is 0.31, and the inverse normalization calculation is 0.31 × 45 N = 13.95 N. After being sorted by timestamp, it is combined with the force values ​​at other times to form a smooth and regular contact signal. This signal truly reproduces the physical interaction force and has no obvious noise fluctuations.

[0038] In step S104, the stiffness change value in the regular contact signal is extracted. If the stiffness change value exceeds a preset stiffness determination threshold, the amplitude and frequency of the regular contact signal are analyzed and compared with a pre-acquired contact type template to identify the dynamic contact type, including: The stiffness change is calculated by the ratio of the force change in the regular contact signal to the synchronous displacement change in the position data. If the stiffness change value exceeds the preset stiffness judgment threshold, then frequency analysis is performed on the regular contact signal to obtain the frequency distribution spectrum. The vibration energy density is calculated based on the frequency distribution spectrum and compared with a pre-acquired contact type template to identify the dynamic contact type.

[0039] It should be noted that when calculating the stiffness change value using the ratio of force change to displacement change in the regularized contact signal, the force change is taken from the difference of force components at consecutive sampling points of the regularized contact signal (which has been denormalized to physical values), and the displacement change is taken from the difference of position data (physical values) that have been time-domain registered in step S102. During calculation, the ratio of the force difference to the displacement difference between two adjacent sampling points (with a time interval of 1ms) is taken to obtain the instantaneous stiffness value. Then, the average is calculated using a sliding window (window size of 10 sampling points) to obtain the stiffness change value (unit: N / m).

[0040] For example, the force value at the previous sampling point is 2.3N, the force value at the current sampling point is 2.8N, the corresponding displacements are 0.02m and 0.021m respectively, the force difference is 0.5N, the displacement difference is 0.001m, the instantaneous stiffness is 500N / m, and the stiffness change value after window averaging is 480N / m.

[0041] In this embodiment, frequency distribution spectrum is obtained by frequency analysis of the regular contact signal, and the stiffness determination threshold is set based on statistical data of stiffness changes in interactions between different materials over the past year. The statistical process follows a standard deviation of 3. Criteria: First, collect a large amount of background stiffness noise data under non-contact and steady-state holding conditions, calculate its probability density distribution, and obtain the mean and standard deviation of stiffness fluctuations; set the basic threshold to the mean plus three times the standard deviation, which can cover most of the range of background noise fluctuations.

[0042] It is worth noting that, in order to meet the high-precision requirements of this invention for subtle tactile perception, the stiffness determination threshold adopts an adaptive threshold floating mechanism based on transient response characteristics. Specifically, the system monitors the second derivative of the rate of change of stiffness (i.e., stiffness acceleration) in real time. This indicator reflects the physical property trend of the contact medium before its numerical value. For example, stiffness exhibits a gradual increase when interacting with soft materials, and a step-like abrupt change when interacting with hard materials.

[0043] Based on this, the system dynamically calculates the stiffness sensitivity factor and fine-tunes the base threshold in real time. When a small and smooth stiffness acceleration is detected, it indicates potential soft material contact, and the threshold is automatically reduced to a specific proportion of the base value (e.g., 0.6 times) to greatly improve the sensitivity to capturing weak deformation signals. When a drastic fluctuation in stiffness acceleration is detected, it indicates potential hard collision or impact, and the threshold is appropriately increased (e.g., 1.2 times) to enhance the ability to suppress transient impact noise. This mechanism ensures that the system can achieve millisecond-level threshold adaptive adjustment in the early stages of dynamic interaction with unknown material properties, significantly improving the accuracy and robustness of contact type identification.

[0044] It should be noted that the frequency analysis uses Fast Fourier Transform (FFT), with a sampling frequency of 1000Hz, 1024 FFT points, a frequency resolution of 0.977Hz, and an analysis range of 0-500Hz (covering the effective frequency range of the tactile signal). The resulting frequency distribution spectrum visually presents the energy distribution of the signal in different frequency bands. For example, a stiffness change of 480N / m exceeds the threshold of 400N / m for metal materials. Performing an FFT on the regular contact signal yields a frequency distribution spectrum showing a significant energy peak in the 150-250Hz frequency band.

[0045] Subsequently, when identifying the dynamic contact type, the vibration energy density is obtained by integrating the power spectral density of each frequency band of the frequency distribution spectrum, and the energy density ratio of the high frequency band (100-500Hz) and the low frequency band (0-100Hz) is calculated respectively.

[0046] It is worth noting that the contact type templates were pre-constructed using supervised feature clustering. The training set collected signal data for different dynamic contact types (puncture, steady-state compression, frictional sliding), covering 10 common materials such as metal, plastic, and rubber, with over 1000 valid samples collected for each type and material. After manual annotation, for each contact type, the energy density ratio data of all samples in the preset high-frequency band (100-500Hz) and low-frequency band (0-100Hz) were extracted.

[0047] Subsequently, the K-Means clustering algorithm was used to calculate the feature center vectors of each type of sample, which were then used as standardized templates for that contact type. The template library formed based on the clustering results has the following characteristics: puncture type is dominated by high frequency (high frequency ≥ 70%, low frequency ≤ 30%), steady-state compression is dominated by low frequency (low frequency ≥ 80%, high frequency ≤ 20%), and frictional sliding is characterized by a balanced energy distribution across frequency bands (both high and low frequency are in the 40%–60% range), thus forming a two-dimensional standardized template library with uniform dimensions.

[0048] Finally, the similarity between the current energy density distribution and each type in the template library is calculated using a cosine similarity algorithm. Specifically, the high-frequency and low-frequency energy density ratios of the current signal are constructed as a two-dimensional feature vector, and the standard energy distributions of each contact type in the template library are constructed as feature vectors of the same dimension. The cosine of the angle between the two vectors is calculated. The closer this value is to 1, the more consistent the two are in their high- and low-frequency energy distribution patterns. The contact type with the highest similarity is the dynamic contact type.

[0049] For example, the current signal has a high-frequency energy density of 68% (0.68) and a low-frequency energy density of 32% (0.32), constructing a vector [0.68, 0.32]. This vector is compared with the puncture type template [0.75, 0.25], showing a high cosine similarity of 0.99; while the similarity with the steady-state compression template [0.15, 0.85] is low. Ultimately, the dynamic contact type is accurately identified as puncture contact.

[0050] In step S105, the process of collecting the dynamic contact parameters of the current interaction, combining the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substituting the multi-dimensional model input vector into a pre-trained friction prediction model to generate a predicted contact force includes: Collect the relative sliding velocity and positive load of the current interaction as contact dynamic parameters; Retrieve the corresponding friction characteristic parameters from the material friction data based on the dynamic contact type; Based on the contact dynamic parameters and the friction characteristic parameters, the thermo-mechanical coupling correction value caused by frictional heat generation is calculated by weighting according to a preset weighting coefficient. The contact dynamic parameters, the friction characteristic parameters, and the thermo-coupling correction values ​​are integrated to form a multi-dimensional model input vector; The input vector of the multidimensional model is substituted into the pre-trained friction prediction model to generate the predicted contact force.

[0051] It should be noted that, firstly, when collecting the relative sliding velocity and positive load of the current interaction as contact dynamic parameters, the relative sliding velocity is collected using a Keyence LK-G80 laser Doppler velocimeter with a sampling frequency of 500Hz, a measurement range of 0-1m / s, and an accuracy of ±0.001m / s. This velocimeter is installed on the side of the robot's end effector to monitor the relative velocity with the interacting object in real time. The positive load uses the Z-axis force component data from the ATI Mini45 six-dimensional force sensor (already collected in S101 and extracted synchronously). After collection, both types of parameters are normalized to minimum and maximum values. The relative sliding velocity is mapped to the [0,1] interval (0m / s corresponds to 0, 1m / s corresponds to 1), and the positive load is mapped to the [0,1] interval (0N corresponds to 0, 50N corresponds to 1), ensuring data scale consistency. For example, when the robot interacts with the rubber material, the laser velocimeter collects a relative sliding speed of 0.3 m / s, which is normalized to 0.3; the force sensor collects a positive load of 15 N, which is normalized to 0.3. Together, they constitute the contact dynamic parameters.

[0052] Next, when retrieving the corresponding frictional characteristic parameters from the material friction data based on the dynamic contact type, the dynamic contact types include three categories: puncture, steady-state extrusion, and frictional sliding (already identified in S104). The material friction data is stored in a JSON database on the local industrial computer and indexed by material type. During the search, the dynamic contact type and material type are used as keywords to extract the corresponding frictional characteristic parameters, including three core parameters: basic friction factor, average surface roughness Ra, and friction temperature coefficient. All parameters have been pre-normalized to the [0,1] interval. For example, if the dynamic contact type is puncture and the material is rubber, the search yields a basic friction factor of 0.32 (normalized to 0.32), an average surface roughness Ra of 0.8 μm (normalized to 0.08), and a friction temperature coefficient of 0.15 (normalized to 0.15), providing basic parameters for thermo-coupling correction.

[0053] Subsequently, based on contact dynamic parameters and friction characteristic parameters, the thermo-mechanical coupling correction value caused by frictional heat generation was calculated using preset weighting coefficients. These weighting coefficients were set as follows: relative sliding velocity 0.4, positive load 0.3, and friction characteristic parameters 0.3. This weighting combination is based on the physical mechanism of frictional heat generation. Relative sliding velocity, being the primary cause of frictional heat generation, is assigned the highest weight; positive load affects the degree of contact tightness, and friction characteristic parameters determine the material's thermal conductivity and the temperature sensitivity of the friction coefficient, all of which are given appropriate weights.

[0054] It should be noted that during the calculation, the normalized values ​​of each parameter are multiplied by their corresponding weights, summed, and then linearly mapped to the correction interval [0, 0.2]. This avoids over-correction that could affect model stability and yields the thermo-coupling correction value. For example, with a relative slip velocity of 0.3 (weight 0.4), a positive load of 0.3 (weight 0.3), and the mean friction characteristic parameter (0.32 + 0.08 + 0.15) / 3 ≈ 0.183 (weight 0.3), the weighted sum is 0.3 × 0.4 + 0.3 × 0.3 + 0.183 × 0.3 ≈ 0.265. After mapping to the correction interval, the thermo-coupling correction value is 0.053.

[0055] When fusing contact dynamic parameters, friction characteristic parameters, and thermo-coupling correction values ​​to form a multi-dimensional model input vector, the parameters are integrated in the following order: relative slip velocity, positive load, basic friction factor, average surface roughness, friction temperature coefficient, and thermo-coupling correction value, forming a 6-dimensional input vector. All components maintain a normalized range of [0,1] to ensure standardized model input. No additional processing is required during the fusion process; parameters are simply concatenated sequentially, preserving the independent characteristics of each parameter while allowing the model to capture the interactions between them. For example, the integrated parameters yield an input vector [0.3,0.3,0.32,0.08,0.15,0.053], fully encompassing the core information of contact dynamics and friction characteristics.

[0056] When the multidimensional model input vector is substituted into the pre-trained friction prediction model to generate predicted contact force, the friction prediction model adopts an LSTM+CNN hybrid architecture, which can not only capture temporal dynamic features, but also extract the spatial correlation between parameters. The model training set is constructed from interactive data of 10 common materials (metal, plastic, rubber, etc.) and 3 dynamic contact types. For each combination, more than 1,000 sets of input vectors and samples corresponding to actual contact forces are collected, totaling more than 30,000 valid samples, which are divided into training set and validation set in an 8:2 ratio.

[0057] In this implementation, the model parameters are set as follows: 2 CNN layers (3×3 kernels, stride 1, max pooling), 1 LSTM layer (64-dimensional hidden layers), and 2 fully connected layers (64-dimensional → 32-dimensional → 1-dimensional). The activation function is ReLU, the loss function is mean squared error (MSE), and the learning rate is initially 0.005, decreasing by 0.1 every 25 epochs until it reaches 0.0005. During training, iteration stops when the validation set loss fluctuation is less than 0.001 for 15 consecutive epochs (maximum training time 200 epochs). After processing, the input vector outputs a predicted value normalized to [0,1]. It is important to note that after the input vector is processed by the model, the predicted contact force sequence, normalized to the [0,1] interval, is directly output, rather than a single numerical value. This is because the LSTM unit in the model has time-series extrapolation capabilities, enabling it to predict the evolution trend of contact force over a future period based on the current input contact dynamic parameters and friction characteristics. This output sequence remains normalized and directly serves as the data source for step S106.

[0058] For example, after the above 6-dimensional input vector is substituted into the model, a normalized prediction sequence containing 100 time steps is output. The sequence value fluctuates between 0.26 and 0.30, representing the trend of the predicted contact force change. This sequence will then be directly used for discretization sampling and deviation feature extraction.

[0059] In step S106, the deviation between the predicted contact force and the actual contact force is calculated. If the deviation exceeds a preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements. A correction feedback signal is then obtained, including: The predicted contact force is discretized and sampled to obtain a time-series friction data sequence. Extract the trend characteristics of the friction data sequence and calculate the deviation between the predicted contact force and the actual contact force; If the deviation does not exceed the preset deviation judgment threshold, the current predicted contact force is used as the correction feedback signal; If the deviation exceeds the deviation determination threshold, the deviation is converted into a deviation feature vector; The weight adjustment value of the corresponding dimension of the friction prediction model is calculated based on the deviation feature vector, the friction characteristic weight is updated, the model output signal is reconstructed and the deviation is checked again until the deviation meets the requirements. The latest model output is then used as the correction feedback signal.

[0060] It should be noted that, firstly, when discretizing the prediction results to obtain the time-series friction data sequence, the discretization sampling extracts the predicted values ​​at 1ms time intervals, and the sampling frequency is consistent with the sensor acquisition frequency. The sequence length is dynamically adjusted according to the interaction duration, using a sliding window mechanism with a window size set to 100 sampling points.

[0061] It is worth noting that when calculating the predicted deviation between the trend characteristics and the actual contact force data, the trend characteristics include the sequence mean, peak value, and slope. The trend characteristics of the predicted contact force are directly derived statistically from the aforementioned time-series friction data; while the trend characteristics of the actual contact force are obtained through parallel statistical calculations on real-time acquired sensor force data (also normalized and extracted within the same time window).

[0062] In this embodiment, the prediction deviation is calculated using the weighted Euclidean distance method. The difference between the three features of the prediction data (mean, peak value, and slope) and the corresponding features of the actual data is calculated, and the difference is summed according to preset weights (such as 0.5, 0.3, and 0.2) to obtain the comprehensive deviation value.

[0063] It should be noted that the weighting is based on the influence of each feature on the fidelity of the tactile feedback. The mean (weight 0.5) represents the steady-state component of the contact force, which directly determines the operator's basic perception of the resistance magnitude, and is therefore given the highest weight. The peak value (weight 0.3) reflects the transient impact intensity, which is related to the identification of the contact boundary and operational safety, and is given the second highest weight. The slope (weight 0.2) characterizes the rate of change of force and contact stiffness, which is a dynamic detail feature and is given a lower weight as an auxiliary indicator. If the overall deviation value exceeds the preset deviation judgment threshold (e.g., 0.05, corresponding to an average error of 2.25N), the model is judged to be inaccurate, and the predicted deviation is transformed into a deviation feature vector. Specifically, the deviation amplitude, rate of change, duration, and polarity are extracted from the predicted deviation and concatenated into a four-dimensional vector, i.e., the deviation feature vector.

[0064] It is important to note that when calculating the weight adjustment value based on the deviation feature vector, a pre-trained weight correction network is used. This network is built on a multilayer perceptron (MLP) architecture and uses a 4-layer fully connected network structure. The input layer has a dimension of 4 (corresponding to the 4 dimensions of the deviation feature vector: deviation magnitude, rate of change, duration, and deviation polarity); the first hidden layer has 64 neurons for extracting the high-dimensional nonlinear mapping of deviation features; the second hidden layer has 32 neurons for feature compression and abstraction; and the output layer has a dimension of 4 (corresponding to the adjustment amount of the 4 friction characteristic weights in the friction prediction model: the weight increment of the basic friction factor, the weight increment of roughness, the weight increment of thermal coupling, and the weight increment of contact dynamic parameters).

[0065] The weights are initialized using a uniform Xavier distribution; the activation function is ReLU (Modified Linear Unit) to accelerate convergence and avoid gradient vanishing; the loss function is mean squared error (MSE), which measures the difference between the adjusted weight values ​​of the network output and the optimal adjusted true values; the optimizer is the Adam algorithm, with an initial learning rate of 0.005 and a learning rate decay strategy, where the learning rate decays by 0.1 every 20 training epochs until it reaches 0.0005 and then stops decaying.

[0066] In this embodiment, the training data comes from an offline-constructed bias-correction database. By simulating the interaction process under different materials (metal, rubber, soft tissue) and contact states in an experimental environment, model errors are artificially introduced (e.g., intentionally biasing the initial weights). The resulting bias feature vectors are recorded, and the optimal weight adjustment values ​​that minimize the bias are solved using gradient descent as labels. The training set contains over 10,000 sample pairs, divided into training and validation sets in an 8:2 ratio.

[0067] The maximum number of iterations is set to 200. The validation set loss is monitored. If the validation set loss decreases by less than 0.001 for 15 consecutive iterations, an early stopping mechanism is triggered to prevent overfitting. The prediction error of the final model is controlled within 0.01.

[0068] Subsequently, the deviation feature vector calculated in real time is input into the trained weight correction network. The network outputs weight adjustment values ​​in four dimensions (the output range is limited to the [-0.1, 0.1] interval by the Tanh activation function to ensure adjustment stability). The update rule is to directly add the original weights to the corresponding weight adjustment values ​​to obtain the updated weights. For example, if the current basic friction factor weight is 0.40 and the network output adjustment value is +0.02, then it is updated to 0.42. The initial weights are not arbitrarily defined, but are taken from the parameters after convergence during the pre-training phase of the friction prediction model.

[0069] Finally, based on the updated friction characteristic weights, the output signal of the friction prediction model is reconstructed to obtain the correction feedback signal. The reconstruction process is as follows: keeping the current input contact dynamic parameter sequence unchanged, only using the updated weights to recalculate the weighted fusion, generating a new contact force prediction sequence containing continuous time steps. This prediction sequence is directly used as a high-fidelity correction feedback signal sequence (maintained in the [0,1] normalization interval) and directly transmitted to the subsequent signal decomposition module.

[0070] In step S107, the step of performing feature decomposition on the correction feedback signal, generating a composite tactile driving signal based on the decomposed high-frequency texture features and low-frequency contact force features, and executing it to obtain a force tactile output consistent with the actual interaction includes: The correction feedback signal is decomposed into high-frequency texture features and low-frequency contact force features by multi-scale decomposition. A micro-vibration driving waveform is generated based on the high-frequency texture features, and a torque resistance command is generated based on the low-frequency contact force features. By superimposing the micro-vibration driving waveform and the torque resistance command, a composite tactile driving signal is constructed. The composite tactile driving signal is converted into a driving electrical signal adapted to the actuator and transmitted to the tactile actuator to drive the tactile actuator to generate physical displacement and reverse resistance, thereby obtaining force tactile output.

[0071] It should be noted that, firstly, when performing multi-scale decomposition on the correction feedback signal to obtain high-frequency texture features and low-frequency contact force features, the Discrete Wavelet Transform (DWT) technique is used, with the Daubechies wavelet (db4) selected as the basis function, and the decomposition scale set to 5 levels. The normalized correction feedback signal sequence output in step S106 is directly used as input. After decomposition, levels 1-2 are high-frequency components, corresponding to 150-350Hz, matching the fine vibration perception range of human fingers, while level 5 is the low-frequency component, corresponding to 0-100Hz, which can reflect the overall resistance trend.

[0072] In this embodiment, the decomposition parameters are calibrated through a training set. The training set collects interaction signals of 10 common materials (metal, rubber, soft tissue, etc.), with 500+ groups for each material, and labels the true distribution of high and low frequency features. During training, the decomposition scale is adjusted to achieve a feature extraction accuracy of 94%, and finally, a 5-level decomposition is determined as the optimal solution.

[0073] For example, in the case of a rubber puncture scenario, the actual force value of the correction feedback signal sequence is about 3.9 N, and the mean of the sequence is about 0.087. After DWT decomposition, the normalized peak value of the extracted high-frequency texture feature sequence is 0.0022 (corresponding to a physical fluctuation of about 0.1 N), and the normalized mean of the low-frequency contact force feature sequence is 0.084 (corresponding to a physical resistance of about 3.8 N), clearly separating the surface micro-texture from the overall macro-resistance information.

[0074] When generating micro-vibration drive waveforms based on high-frequency texture features and torque resistance commands based on low-frequency contact force features, the high-frequency texture features first calculate the energy density and map it to the micro-vibration amplitude (range 0.01-0.1mm) according to the energy ratio. The frequency is fixed at 200-300Hz (the range sensitive to human touch). The waveform uses sinusoidal modulation to ensure smooth vibration without impact. The low-frequency contact force features are denormalized to [0,45]N, and commands are generated according to the linear mapping relationship between force and torque (mapping coefficient 0.02N·m / N, i.e., 1N force corresponds to 0.02N·m torque). The command output delay is ≤5ms.

[0075] It should be noted that the determination of this mapping relationship is based on the kinematic parameters of the haptic actuator and the force scaling strategy: First, a "remote-local" force scaling ratio is set (e.g., 1:3.8) to scale the large load force from a distance to a comfortable range for the human hand to prevent fatigue; second, combined with the effective lever arm length of the haptic handle (e.g., 0.076m), the scaled linear force is converted into the output torque required by the motor. This coefficient 0.02 is the product of the scaling ratio and the lever arm length. This ensures the accuracy of physical quantity conversion.

[0076] For example, the high-frequency texture feature has an energy density of 0.01 N² / Hz, generating a sinusoidal vibration waveform with an amplitude of 0.08 mm and a frequency of 220 Hz; the low-frequency contact force feature has an energy density of 3.8 N, which, after scaling and lever arm conversion, generates a torque resistance command of 0.076 N·m, corresponding to a continuous reverse resistance of 1.0 N at the effective lever arm of the handle.

[0077] Subsequently, when constructing the composite tactile drive signal by superimposing the micro-vibration drive waveform and the torque resistance command, a linear fusion method is adopted for superposition. An amplitude modulation coefficient of 0.7 is applied to the high-frequency waveform to avoid excessive vibration masking the low-frequency resistance sensing. Before fusion, the two types of signals are synchronized in time (interpolated to align the high-frequency waveform based on the low-frequency command timestamp). After synchronization, the signal sampling frequency is maintained at 1000Hz, consistent with the sensor acquisition frequency. The composite signal is finally normalized to the [0,1] interval to adapt to the actuator drive voltage range (0-5V).

[0078] For example, the peak value of the micro-vibration waveform is 0.08mm, the torque resistance command corresponds to 1.0N resistance, and the composite signal after superposition is based on the 1.0N resistance with periodic micro-perturbations, which not only retains the overall contact feel, but also restores the subtle texture feel of the rubber surface.

[0079] Finally, when the tactile actuator is driven by a composite tactile drive signal to generate physical displacement and reverse resistance to obtain force tactile output, the tactile actuator uses a voice coil motor (model: SMACLCA25) with a flexible linkage structure. The motor has a maximum stroke of 0.5mm, a response frequency of 1kHz, and a thrust range of 0-5N, meeting the requirements for micro-vibration and resistance simulation. The drive signal is input to the motor controller after D / A conversion (16-bit accuracy). The controller adjusts the motor displacement and thrust according to the signal instructions. The physical displacement corresponds to the micro-vibration waveform (maximum 0.1mm), and the reverse resistance corresponds to the stress torque resistance instruction (maximum 5N).

[0080] For example, after the composite signal is input, the motor outputs a 220Hz micro-vibration of 0.08mm, while providing a continuous reverse resistance of 1.0N. When the operator holds the handle, they can feel the vibration of the rough texture of the rubber surface and also sense the stable resistance during puncture, which is highly consistent with the actual interactive tactile sensation of the remote robot.

[0081] In summary, this invention discloses a remote robot interaction method based on tactile feedback, comprising: acquiring force data, position data, and material friction data of the remote robot interacting with the environment; obtaining an initial contact signal through temporal registration and differential processing, and extracting force components after correction to generate a regular contact signal; identifying the dynamic contact type using a hyperstiffness threshold, constructing a multi-dimensional input vector by combining contact dynamic parameters and friction data, and substituting it into a pre-trained model to generate a predicted contact force; adjusting the friction characteristic weights using a hyperdeviation threshold to obtain a correction feedback signal; and decomposing high-frequency texture and low-frequency contact force features to generate a composite tactile driving signal, which drives the actuator to achieve high-fidelity force-tactile output. This method achieves high-fidelity force-tactile output and high-fidelity force-tactile feedback for remote robot interaction, meeting the stringent requirements for accuracy and safety in remote operations.

[0082] Reference Figure 2 The second embodiment of the present invention provides a robot remote interaction system based on haptic feedback, comprising: The data acquisition module is used to acquire force data, position data, and material friction data of the interaction between the remote robot and the environment; The contact signal module is used to perform time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal. The signal processing module is used to construct a state vector based on the initial contact signal, and substitute it into a preset state transition matrix to obtain a predicted state vector. It also collects real-time force observations and corrects the predicted state vector based on the observations, extracts the force components of the corrected state vector, and obtains a regular contact signal. The contact judgment module is used to extract the stiffness change value in the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, the amplitude and frequency of the regular contact signal are analyzed and compared with the pre-acquired contact type template to identify the dynamic contact type. The contact force prediction module collects the dynamic contact parameters of the current interaction, combines the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitutes the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force. The signal correction module calculates the deviation between the predicted contact force and the actual contact force. If the deviation exceeds a preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements, and then a correction feedback signal is obtained. The tactile output module performs feature decomposition on the correction feedback signal, generates a composite tactile drive signal based on the decomposed high-frequency texture features and low-frequency contact force features, and executes it to obtain a force tactile output consistent with the actual interaction.

[0083] It should be noted that the tactile feedback-based robot remote interaction system provided in this embodiment of the invention is used to execute all the process steps of the tactile feedback-based robot remote interaction method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0084] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0085] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for remote robot interaction based on haptic feedback, characterized in that, include: Acquire force data, position data, and material friction data of the interaction between the remote robot and its environment; The force data and the position data are time-domain registered and differentiated to obtain the initial contact signal; A state vector is constructed based on the initial contact signal, and a predicted state vector is obtained by substituting it into a preset state transition matrix. Real-time force observations are collected, and the predicted state vector is corrected based on the observations. The force components of the corrected state vector are extracted to obtain a regular contact signal. Extract the stiffness change value from the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, analyze the amplitude and frequency of the regular contact signal and compare it with the pre-acquired contact type template to identify the dynamic contact type. Collect the dynamic parameters of the current interaction, combine the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitute the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force; The deviation between the predicted contact force and the actual contact force is calculated. If the deviation exceeds the preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements, and then a correction feedback signal is obtained. The correction feedback signal is decomposed into features. Based on the decomposed high-frequency texture features and low-frequency contact force features, a composite tactile driving signal is generated and executed to obtain a force tactile output consistent with the actual interaction.

2. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The acquisition of force data, position data, and material friction data of the remote robot's interaction with the environment includes: Force data of the interaction between the remote robot and its environment is collected using force sensors; The position data of the remote robot is collected using a position encoder; Retrieve pre-stored material friction data, which includes basic friction factors and surface texture information for different materials.

3. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The step of performing time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal includes: Align the force data and the position data according to the timestamp to generate a synchronization data frame; The synchronous data frame is differentiated to calculate the gradient of force change; If the force change gradient exceeds a preset gradient determination threshold, the corresponding synchronous data frame is locked as instantaneous contact data. The instantaneous contact data is extracted and combined with the position data to generate an initial contact signal.

4. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The process of constructing a state vector based on the initial contact signal, substituting it into a preset state transition matrix to obtain a predicted state vector, acquiring real-time force observations and correcting the predicted state vector based on the observations, extracting the force components of the corrected state vector, and obtaining a regularized contact signal includes: A state vector is constructed based on the initial contact signal, and the predicted state vector is obtained by substituting it into a preset state transition matrix. Collect real-time force observations, extract the timestamps of the observations and the extrapolation timestamps of the predicted state vectors, calculate the time difference, and generate a delay compensation factor based on the time difference; The observed values ​​are corrected according to the delay compensation factor to obtain aligned observation data; Based on a preset noise suppression coefficient, the aligned observation data and the predicted state vector are weighted and fused, and sensor noise is filtered out through signal smoothing to obtain a corrected state vector. Force component values ​​are extracted from the corrected state vector to obtain a regular contact signal.

5. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The process involves extracting the stiffness variation value from the regular contact signal. If the stiffness variation value exceeds a preset stiffness judgment threshold, the amplitude and frequency of the regular contact signal are analyzed and compared with a pre-acquired contact type template to identify the dynamic contact type, including: The stiffness change is calculated by the ratio of the force change in the regular contact signal to the synchronous displacement change in the position data. If the stiffness change value exceeds the preset stiffness judgment threshold, then frequency analysis is performed on the regular contact signal to obtain the frequency distribution spectrum. The vibration energy density is calculated based on the frequency distribution spectrum and compared with a pre-acquired contact type template to identify the dynamic contact type.

6. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The process involves collecting dynamic contact parameters from the current interaction, combining them with the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and then substituting this multi-dimensional model input vector into a pre-trained friction prediction model to generate a predicted contact force, including: Collect the relative sliding velocity and positive load of the current interaction as contact dynamic parameters; Retrieve the corresponding friction characteristic parameters from the material friction data based on the dynamic contact type; Based on the contact dynamic parameters and the friction characteristic parameters, the thermo-mechanical coupling correction value caused by frictional heat generation is calculated by weighting according to a preset weighting coefficient. The contact dynamic parameters, the friction characteristic parameters, and the thermo-coupling correction values ​​are integrated to form a multi-dimensional model input vector; The input vector of the multidimensional model is substituted into the pre-trained friction prediction model to generate the predicted contact force.

7. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The process involves calculating the deviation between the predicted contact force and the actual contact force. If the deviation exceeds a preset deviation threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements. A correction feedback signal is then obtained, including: The predicted contact force is discretized and sampled to obtain a time-series friction data sequence. Extract the trend characteristics of the friction data sequence and calculate the deviation between the predicted contact force and the actual contact force; If the deviation does not exceed the preset deviation judgment threshold, the current predicted contact force is used as the correction feedback signal; If the deviation exceeds the deviation determination threshold, the deviation is converted into a deviation feature vector; The weight adjustment value of the corresponding dimension of the friction prediction model is calculated based on the deviation feature vector, the friction characteristic weight is updated, the model output signal is reconstructed and the deviation is checked again until the deviation meets the requirements. The latest model output is then used as the correction feedback signal.

8. The robot remote interaction method based on haptic feedback according to claim 1, characterized in that, The step of performing feature decomposition on the correction feedback signal, generating a composite tactile driving signal based on the decomposed high-frequency texture features and low-frequency contact force features, and executing it to obtain a force-tactile output consistent with actual interaction includes: The correction feedback signal is decomposed into high-frequency texture features and low-frequency contact force features by multi-scale decomposition. A micro-vibration driving waveform is generated based on the high-frequency texture features, and a torque resistance command is generated based on the low-frequency contact force features. By superimposing the micro-vibration driving waveform and the torque resistance command, a composite tactile driving signal is constructed. The composite tactile driving signal is converted into a driving electrical signal adapted to the actuator and transmitted to the tactile actuator to drive the tactile actuator to generate physical displacement and reverse resistance, thereby obtaining force tactile output.

9. A robot remote interaction system based on haptic feedback, characterized in that, include: The data acquisition module is used to acquire force data, position data, and material friction data of the interaction between the remote robot and the environment; The contact signal module is used to perform time-domain registration and differentiation processing on the force data and the position data to obtain the initial contact signal. The signal processing module is used to construct a state vector based on the initial contact signal, and substitute it into a preset state transition matrix to obtain a predicted state vector. It also collects real-time force observations and corrects the predicted state vector based on the observations, extracts the force components of the corrected state vector, and obtains a regular contact signal. The contact judgment module is used to extract the stiffness change value in the regular contact signal. If the stiffness change value exceeds the preset stiffness judgment threshold, the amplitude and frequency of the regular contact signal are analyzed and compared with the pre-acquired contact type template to identify the dynamic contact type. The contact force prediction module collects the dynamic contact parameters of the current interaction, combines the dynamic contact type and the material friction data to construct a multi-dimensional model input vector, and substitutes the multi-dimensional model input vector into the pre-trained friction prediction model to generate the predicted contact force. The signal correction module calculates the deviation between the predicted contact force and the actual contact force. If the deviation exceeds a preset deviation judgment threshold, the friction characteristic weights of the friction prediction model are adjusted, the predicted contact force is regenerated, and the deviation is checked again until the deviation meets the requirements, and then a correction feedback signal is obtained. The tactile output module performs feature decomposition on the correction feedback signal, generates a composite tactile drive signal based on the decomposed high-frequency texture features and low-frequency contact force features, and executes it to obtain a force tactile output consistent with the actual interaction.

Images