Robotic grasping method and apparatus based on tactile perception

Abstract

The application provides a robot grasping method and device based on tactile perception. The method comprises the following steps: determining a smooth pressure signal and a target key feature set according to high-dimensional tactile data; determining an adjustment angle of a mechanical claw according to a humidity distribution feature and a first mapping table; determining a viscous compensation grasping pressure threshold according to the adjustment angle and a viscous stretching morphology feature in combination with a second mapping table; determining a force feedback deviation, a baseline closure angle of the mechanical claw and a baseline grasping pressure according to the grasping pressure threshold, a debris embedding area feature and a current grasping pressure; and determining a target control parameter of the mechanical claw and controlling the mechanical claw to grasp the wet garbage in combination with a pre-constructed deep reinforcement learning grasping strategy network. In this way, the control parameter of the mechanical claw is determined by combining the humidity interference, the viscous stretching and the debris embedding area feature of the wet garbage, the control parameter of the mechanical claw can be adjusted in real time according to the characteristics of the wet garbage, and the accuracy of the wet garbage grasping is improved.

Description

Technical Field

[0001] This invention relates to the field of waste collection technology, and more specifically to a robot grasping method and device based on tactile perception. Background Technology

[0002] As a crucial component of smart city construction, sanitation robots' grasping technology directly impacts the efficiency and quality of waste sorting and treatment, making it indispensable. With accelerating urbanization and increasing environmental demands, enabling robots to achieve precise and efficient grasping in complex and ever-changing waste environments has become a core issue driving the development of smart sanitation. Traditional sanitation robots rely heavily on visual recognition and fixed grasping strategies, which can easily lead to grasping failures or damage to the target. Existing sanitation robots struggle to reliably handle complex waste types in real-world wet waste environments, characterized by varying hardness, smoothness, fragility, and the interference of moisture and stickiness, becoming a key technological bottleneck restricting their engineering applications. Summary of the Invention

[0003] This application aims to provide a robot grasping method and device based on tactile perception to improve the accuracy of grasping wet waste.

[0004] Firstly, a robot grasping method based on tactile perception is provided, the method comprising:

[0005] High-dimensional tactile data is acquired, and the pressure distribution in the high-dimensional tactile data is denoised to obtain a smooth pressure signal.

[0006] Based on the smooth pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features.

[0007] The adjustment angle of the mechanical gripper is determined based on the humidity distribution characteristics and the first mapping table, whereby the first mapping table represents the mapping relationship between humidity and the angle of the mechanical gripper.

[0008] The stickiness of the waste is determined based on the adjustment angle and the viscous stretching morphology characteristics. The gripping pressure threshold after stickiness compensation is determined by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold.

[0009] The force feedback deviation is determined based on the gripping pressure threshold after viscosity compensation, the characteristics of the debris embedding area, and the current gripping pressure.

[0010] In response to the force feedback deviation being greater than a preset deviation threshold, the baseline closing angle and baseline gripping pressure of the mechanical gripper are determined based on the force feedback deviation.

[0011] Based on the baseline closure angle, the baseline grasping pressure, and the pre-constructed deep reinforcement learning grasping policy network, the target control parameters of the robotic gripper are determined, including the target closure angle and the target grasping pressure.

[0012] The robotic gripper is controlled to grasp wet waste according to the target control parameters.

[0013] Optionally, based on the smoothed pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, including:

[0014] The smoothed pressure signal and its spatial neighborhood are input into a lightweight convolutional neural network to obtain an initial key feature set.

[0015] By combining a preset screening threshold with the analysis method of support vector machine, the features in the initial key feature set are screened, and the regions with excessive humidity and local mutation regions caused by debris embedding are extracted to determine the target key feature set.

[0016] Optionally, determining the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table includes:

[0017] The humidity distribution characteristics are quantitatively evaluated to obtain the degree of humidity interference;

[0018] If the humidity interference level is greater than the preset humidity interference threshold, the adjustment angle of the mechanical gripper is determined according to the first mapping table.

[0019] Optionally, the force feedback deviation is determined based on the viscosity-compensated gripping pressure threshold, the characteristics of the debris embedding region, and the current gripping pressure, including:

[0020] The gripping pressure threshold after viscosity compensation is corrected based on the characteristics of the debris embedding region to obtain the debris-sensitive reference pressure.

[0021] The difference between the current gripping pressure and the debris-sensitive reference pressure is defined as the force feedback deviation.

[0022] Optionally, in response to the force feedback deviation being greater than a preset deviation threshold, determining the baseline closing angle and baseline gripping pressure of the robotic gripper based on the force feedback deviation includes:

[0023] The optimal parameter adjustment instructions for the mechanical gripper are determined based on the magnitude and direction of the force feedback deviation. The optimal parameter adjustment instructions include angle adjustment instructions and gripping pressure adjustment instructions.

[0024] During the process of the robotic gripper dynamically correcting itself according to the optimized parameter adjustment command, the stable reference values ​​of the robotic gripper's angle and gripping pressure are determined as the baseline closing angle and baseline gripping pressure.

[0025] Optionally, the target control parameters of the robotic gripper are determined based on the baseline closure angle, the baseline gripping pressure, and the pre-constructed deep reinforcement learning gripping policy network, including:

[0026] The state vector is input into a pre-constructed deep reinforcement learning grasping policy network, and the output is the closing angle reinforcement adjustment amount and the grasping pressure reinforcement adjustment amount. The state vector includes at least the humidity interference threshold, viscosity degree, force feedback contrast, debris embedding region features, baseline closing angle and baseline grasping pressure.

[0027] The target closure angle and the target gripping pressure are determined based on the closure angle enhancement adjustment amount, the gripping pressure enhancement adjustment amount, the baseline closure angle, and the baseline gripping pressure.

[0028] Optionally, determining the target closure angle and the target gripping pressure based on the closure angle enhancement adjustment amount, the gripping pressure enhancement adjustment amount, the baseline closure angle, and the baseline gripping pressure includes:

[0029] Obtain the actual latency of the robot, and determine the dimensionless latency evaluation index based on the actual latency;

[0030] When the dimensionless time delay evaluation index is less than or equal to the preset time delay evaluation threshold, the closing angle enhancement adjustment amount is superimposed on the baseline closing angle, and the gripping pressure enhancement adjustment amount is superimposed on the baseline gripping pressure to obtain the target closing angle and the target gripping pressure.

[0031] If the dimensionless time delay evaluation index is greater than the preset time delay evaluation threshold, the baseline closure angle and the baseline grasping pressure are determined as the target closure angle and the target grasping pressure.

[0032] Alternatively, the crawling method may also include:

[0033] During the grasping process of the robotic gripper, multi-dimensional grasping feedback data is acquired in real time. The multi-dimensional grasping feedback data includes at least pressure, contact area, and sliding trend.

[0034] A set of multi-frame crawling parameter vectors is constructed based on the multi-dimensional crawling feedback data;

[0035] The stability evaluation index of the crawling parameters is determined based on the set of multi-frame crawling parameter vectors;

[0036] If the stability evaluation index of the grasping parameters is greater than the preset stability threshold, the deep reinforcement learning grasping strategy network is given a low reward.

[0037] Secondly, a robotic grasping device based on tactile perception is provided, the device comprising:

[0038] The acquisition module is used to acquire high-dimensional tactile data and perform noise reduction processing on the pressure distribution in the high-dimensional tactile data to obtain a smooth pressure signal.

[0039] The extraction module is used to extract features from the high-dimensional tactile data based on the smooth pressure signal to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features.

[0040] The first determining module is used to determine the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table, wherein the first mapping table represents the mapping relationship between humidity and mechanical gripper angle.

[0041] The second determining module is used to determine the stickiness of the waste based on the adjustment angle and the sticky stretching morphological characteristics, and to determine the gripping pressure threshold after stickiness compensation by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold.

[0042] The third determining module is used to determine the force feedback deviation based on the gripping pressure threshold after viscosity compensation, the characteristics of the debris embedding area, and the current gripping pressure.

[0043] The fourth determining module is used to determine the baseline closing angle and baseline gripping pressure of the mechanical gripper based on the force feedback deviation in response to the force feedback deviation being greater than a preset deviation threshold.

[0044] The fifth determining module is used to determine the target control parameters of the robotic gripper based on the baseline closing angle, the baseline gripping pressure, and the pre-constructed deep reinforcement learning gripping strategy network. The target control parameters include the target closing angle and the target gripping pressure.

[0045] The control module is used to control the mechanical claw to grasp wet waste according to the target control parameters.

[0046] Thirdly, an electronic device is provided, comprising:

[0047] The memory is configured to store instructions;

[0048] The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the tactile perception-based robot grasping method provided in the first aspect of the embodiments of this application.

[0049] Based on the aforementioned tactile perception-based robotic grasping method, high-dimensional tactile data is acquired, and the pressure distribution in the high-dimensional tactile data is denoised to obtain a smooth pressure signal. Based on the smooth pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features. The adjustment angle of the robotic gripper is determined according to the humidity distribution features and a first mapping table, which represents the mapping relationship between humidity and the robotic gripper angle. The stickiness of the waste is determined based on the adjustment angle and the viscous stretching morphology features, and the stickiness is further determined by combining this with a second mapping table. The compensated gripping pressure threshold is represented by a second mapping table, which characterizes the mapping relationship between viscosity and gripping pressure threshold. Force feedback deviation is determined based on the viscosity-compensated gripping pressure threshold, debris embedding region characteristics, and current gripping pressure. In response to a force feedback deviation exceeding a preset deviation threshold, the baseline closing angle and baseline gripping pressure of the robotic gripper are determined based on the force feedback deviation. Target control parameters for the robotic gripper are determined based on the baseline closing angle, baseline gripping pressure, and a pre-constructed deep reinforcement learning gripping strategy network. These target control parameters include the target closing angle and target gripping pressure. The robotic gripper is then controlled to grip wet waste according to these target control parameters. The beneficial effect of this application is that by performing multi-dimensional tactile perception on wet waste, and combining the wet waste's humidity interference, viscous stretching, and debris embedding region characteristics to determine the robotic gripper's control parameters, the control parameters can be adjusted in real time according to the characteristics of wet waste, thereby improving the accuracy of wet waste gripping. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the robot grasping method based on tactile perception provided in the embodiments of this application;

[0051] Figure 2 This is a schematic diagram of the structure of the robot grasping device based on tactile perception provided in the embodiments of this application;

[0052] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0053] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0054] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0055] The robot grasping method and apparatus based on tactile perception provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0056] Please see Figure 1 This is a flowchart illustrating a robot grasping method based on tactile perception provided in an embodiment of this application. This method is applied to electronic devices. Figure 1 As shown, the capture method includes the following steps S100 to S800.

[0057] Step S100: Acquire high-dimensional tactile data and perform noise reduction processing on the pressure distribution in the high-dimensional tactile data to obtain a smooth pressure signal.

[0058] In this embodiment, high-dimensional tactile data can be understood as data collected by a multi-dimensional tactile sensor array disposed on the gripping part of a robotic claw. The multi-dimensional tactile sensor array may include, but is not limited to, pressure array sensors and shear force sensors. High-dimensional tactile data may include, but is not limited to, pressure, changes in contact area, and dynamic responses over time. The robotic claw may be the robotic claw of various special-purpose robots or a robotic arm with gripping capabilities. To address the humidity interference and viscous stretching characteristics of wet filter droplets, a pre-set wet waste-specific filter is first used to denoise the pressure distribution in the high-dimensional tactile data to suppress large-scale trend drift and viscous tail distortion, obtaining an initial processed pressure signal. In one example, the wet waste-specific filter may include a high-pass or detrending filter unit for suppressing low-frequency drift caused by humidity, and a hysteresis compensation unit for compensating for tail distortion caused by viscous stretching.

[0059] After obtaining the initial processed pressure signal, wavelet denoising and time-series smoothing are performed to obtain a smoothed pressure signal for feature extraction. First, a pressure distribution matrix is ​​extracted from the original high-dimensional data. Combined with a preset humidity threshold, data segments significantly affected by temperature interference are initially screened. For the screened pressure distribution, a lightweight convolutional neural network (CNN) is used to extract local spatial features, identify abnormal patterns caused by viscous stretching, and output optimized pressure features. In one example, the lightweight CNN employs a structure combining depthwise separable convolution, channel pruning, and parameter quantization to reduce network inference latency while maintaining accuracy in extracting features related to humidity distribution, viscous stretching morphology, and debris embedding regions. For residual high-frequency noise and local distortion, wavelet transform is further used to decompose the optimized pressure features, separating noise components from effective pressure components to obtain a refined pressure signal. If the variance of the refined pressure signal is still higher than the preset smoothing threshold, wavelet denoising is repeated until the smoothing requirement is met. Finally, time series features are extracted from the refining pressure signal. Combined with the slow change characteristics of pressure over time in the wet waste scenario, the time series is processed by mean filtering or sliding window weighted averaging to obtain the final stable and smooth pressure signal.

[0060] Step S200: Based on the smoothed pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features.

[0061] In this embodiment, when extracting features from high-dimensional tactile data based on smoothed pressure signals, the smoothed pressure signal and its spatial neighborhood can be input into a lightweight convolutional neural network for feature extraction, resulting in an initial key feature set containing initial humidity distribution features, initial viscous stretching morphology features, and initial debris embedding region features. Then, by combining a preset threshold with classification or regression analysis using support vector machines, components reflecting signal amplitude fluctuations and local anomalies in the initial key feature set are filtered to extract humidity-exceeding regions and local abrupt change regions caused by debris embedding, thus obtaining the target key feature set.

[0062] Step S300: Determine the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table, wherein the first mapping table represents the mapping relationship between humidity and mechanical gripper angle.

[0063] In this embodiment, humidity distribution features are extracted from the target key feature set, and the humidity distribution features are quantitatively evaluated to obtain the degree of humidity interference. When the degree of humidity interference exceeds a preset humidity interference threshold, a first mapping table pre-established based on experimental data is invoked to calculate the adjustment angle of the mechanical gripper, i.e., the closing angle adjustment amount. The mechanical gripper angle is the closing angle of the mechanical gripper. The first mapping table represents the mechanical gripper closing angle values ​​corresponding to different humidity ranges; the higher the humidity, the smaller the mechanical gripper closing angle. In a specific example, when constructing the first mapping table, mechanical gripper grasping test data of multiple sets of garbage samples are collected under different humidity conditions, and the corresponding closing angle and grasping success rate are recorded. The relationship between the humidity interference index and the closing angle adjustment amount with the highest grasping success rate is fitted to form a one-to-one correspondence between humidity and the mechanical gripper closing angle.

[0064] Step S400: Determine the stickiness of the waste based on the adjustment angle and the sticky stretching morphology characteristics, and determine the gripping pressure threshold after stickiness compensation by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold.

[0065] In this embodiment, based on the determined adjustment angle of the mechanical gripper, a support vector machine can be used to determine the degree of stickiness of the waste to address the contact area distortion caused by viscous stretching. A second mapping table based on experimental data is then used to obtain the gripping pressure threshold after viscosity compensation. In a specific example, when constructing the second mapping table, gripping tests are conducted under different viscosity conditions. Whether the waste slips or breaks under different gripping pressures is recorded. Under the condition of ensuring no breakage or slippage, the maximum non-destructive gripping pressure corresponding to each viscosity level is determined. The relationship between the degree of stickiness and the maximum non-destructive gripping pressure is then fitted to form a mapping relationship between the degree of stickiness and the gripping pressure threshold.

[0066] Specifically, data such as contact area and pressure increase rate are used as inputs to a Support Vector Machine (SVM). The SVM outputs the corresponding viscosity level (which can be considered a dimensionless level or interval label). A mapping is established based on a pre-constructed second mapping table. .in, This indicates the gripping pressure threshold (in Pa or N). Indicating the degree of viscosity, the function The data is fitted from experimental data. Through this mapping, the upper limit of the allowable pressure is automatically reduced when the viscosity is high, resulting in a viscosity-compensated gripping pressure threshold, which is used for subsequent gripping pressure control and compensation.

[0067] Step S500: Determine the force feedback deviation based on the gripping pressure threshold after viscosity compensation, the characteristics of the debris embedding area, and the current gripping pressure.

[0068] Step S600: In response to the force feedback deviation being greater than a preset deviation threshold, determine the baseline closing angle and baseline gripping pressure of the mechanical gripper based on the force feedback deviation.

[0069] In this embodiment, the gripping pressure threshold after viscosity compensation is fused with the debris embedding characteristics and the current gripping pressure to calculate the force feedback deviation. If the absolute value of the force feedback deviation exceeds a preset deviation threshold, a preset compensation algorithm is activated to dynamically adjust the gripping pressure based on the magnitude and direction of the deviation, limiting the gripping pressure to the range of 0 to the gripping pressure threshold. This generates optimized adjustment instructions for the robotic gripper that comprehensively consider the effects of humidity, viscosity, and debris embedding. These optimized parameter adjustment instructions may include, but are not limited to, angle adjustment instructions and gripping pressure adjustment instructions. During the dynamic correction process of the robotic gripper according to the optimized parameter adjustment instructions, the stable reference values ​​of the robotic gripper's angle and gripping pressure are determined as the baseline closing angle and baseline gripping pressure.

[0070] Step S700: Determine the target control parameters of the robotic gripper based on the baseline closure angle, the baseline gripping pressure, and the pre-constructed deep reinforcement learning gripping policy network. The target control parameters include the target closure angle and the target gripping pressure.

[0071] In this embodiment, a pre-constructed deep reinforcement learning grasping policy network is first obtained. A state vector, composed of humidity interference threshold, viscosity level, force feedback contrast, debris embedding region features, baseline closure angle, and baseline grasping pressure, is input into the pre-constructed deep reinforcement learning grasping policy network, which then outputs a closure angle reinforcement adjustment and a grasping pressure reinforcement adjustment. The target closure angle and target grasping pressure are then determined based on the closure angle reinforcement adjustment, the grasping pressure reinforcement adjustment, the baseline closure angle, and the baseline grasping pressure.

[0072] Step S800: Control the mechanical claw to grab wet waste according to the target control parameters.

[0073] Through steps S100-S800, high-dimensional tactile data is acquired, and the pressure distribution in the high-dimensional tactile data is denoised to obtain a smooth pressure signal. Based on the smooth pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features. The adjustment angle of the mechanical claw is determined according to the humidity distribution features and a first mapping table, which represents the mapping relationship between humidity and the angle of the mechanical claw. The stickiness of the waste is determined according to the adjustment angle and the viscous stretching morphology features, and the stickiness compensation is determined in conjunction with a second mapping table. The compensation-adjusted gripping pressure threshold is used, and a second mapping table represents the mapping relationship between the viscosity level and the gripping pressure threshold. The force feedback deviation is determined based on the viscosity-compensated gripping pressure threshold, the characteristics of the debris embedding region, and the current gripping pressure. In response to a force feedback deviation exceeding a preset deviation threshold, the baseline closing angle and baseline gripping pressure of the robotic gripper are determined based on the force feedback deviation. The target control parameters of the robotic gripper, including the target closing angle and target gripping pressure, are determined based on the baseline closing angle, baseline gripping pressure, and a pre-constructed deep reinforcement learning gripping strategy network. The robotic gripper is then controlled to grip wet waste according to these target control parameters. Thus, by performing multi-dimensional tactile perception of wet waste and combining the wet waste's humidity interference, viscous stretching, and debris embedding region characteristics to determine the robotic gripper's control parameters, the control parameters can be adjusted in real time according to the characteristics of wet waste, thereby improving the accuracy of wet waste gripping.

[0074] In some implementations, the step of extracting features from the high-dimensional tactile data based on the smoothed pressure signal to obtain a target key feature set includes:

[0075] The smoothed pressure signal and its spatial neighborhood are input into a lightweight convolutional neural network to obtain an initial key feature set.

[0076] By combining a preset screening threshold with the analysis method of support vector machine, the features in the initial key feature set are screened, and the regions with excessive humidity and local mutation regions caused by debris embedding are extracted to determine the target key feature set.

[0077] Specifically, during feature extraction, the smoothed pressure signal and its spatial neighborhood are used as input to a lightweight convolutional neural network to extract multi-scale features related to humidity distribution, viscous stretching morphology, and debris embedding regions, resulting in an initial key feature set. Support vector machine (SVM) analysis methods can include classification or regression analysis. For components in the initial key feature set reflecting signal amplitude fluctuations and local anomalies, SVM is introduced for regression or classification analysis to distinguish normal fluctuations from anomalies caused by humidity / debris, resulting in a feature enhancement signal. The feature enhancement signal is then filtered using a preset threshold to extract humidity-exceeding regions and local abrupt change regions caused by debris embedding, forming an abnormal distribution signal. For regions in the abnormal distribution signal significantly affected by humidity, a lightweight convolutional neural network is used again for secondary discrimination to refine the intensity pattern of debris embedding. Gaussian filtering is then used to smooth the feature enhancement signal, suppressing residual noise and edge fluctuations, ultimately yielding the target key feature set for subsequent parameter mapping.

[0078] In some embodiments, determining the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table includes:

[0079] The humidity distribution characteristics are quantitatively evaluated to obtain the degree of humidity interference;

[0080] If the humidity interference level is greater than the preset humidity interference threshold, the adjustment angle of the mechanical gripper is determined according to the first mapping table.

[0081] Specifically, humidity distribution features are extracted from the target key feature set, and these features are quantitatively evaluated to obtain the degree of humidity interference. When the degree of humidity interference exceeds a preset humidity interference threshold, a first mapping table established based on extensive experimental data is invoked to obtain the adjustment angle of the robotic gripper. The mapping relationship of the first mapping table can be abstractly represented as: ,in, To adjust the angle of the mechanical gripper, For dimensionless or standardized humidity indices, the function The results were obtained through experimental fitting. The above mapping relationship ensures that the closing angle of the mechanical gripper is appropriately reduced when humidity increases, avoiding excessive compression of wet waste. When necessary, a lightweight convolutional neural network or support vector machine can be used to perform regression correction on the initial adjustment results of the mechanical gripper's angle, ensuring that the closing angle adjustment meets both anti-breakage requirements and gripping stability.

[0082] In some implementations, determining the force feedback deviation based on the viscosity-compensated gripping pressure threshold, the characteristics of the debris embedding region, and the current gripping pressure includes:

[0083] The gripping pressure threshold after viscosity compensation is corrected based on the characteristics of the debris embedding region to obtain the debris-sensitive reference pressure.

[0084] The difference between the current gripping pressure and the debris-sensitive reference pressure is defined as the force feedback deviation.

[0085] Specifically, the gripping pressure threshold after viscosity compensation is the basic safe pressure upper limit that already takes into account the viscosity of the waste. Since debris embedding reduces the pressure bearing capacity of the gripping contact surface—harder debris is prone to causing localized stress concentration, while softer debris is easily compacted, leading to pressure loss—the gripping pressure threshold after viscosity compensation needs to be negatively corrected based on the characteristics of the debris embedding area to obtain a debris-sensitive reference pressure. The difference between the current gripping pressure and the debris-sensitive reference pressure is then defined as the force feedback deviation.

[0086] In some implementations, the step of determining the baseline closing angle and baseline gripping pressure of the robotic gripper based on the force feedback deviation in response to the force feedback deviation being greater than a preset deviation threshold includes:

[0087] The optimal parameter adjustment instructions for the mechanical gripper are determined based on the magnitude and direction of the force feedback deviation. The optimal parameter adjustment instructions include angle adjustment instructions and gripping pressure adjustment instructions.

[0088] During the process of the robotic gripper dynamically correcting itself according to the optimized parameter adjustment command, the stable reference values ​​of the robotic gripper's angle and gripping pressure are determined as the baseline closing angle and baseline gripping pressure.

[0089] Specifically, when the force feedback deviation exceeds a preset deviation threshold, a compensation algorithm is triggered. Based on the magnitude and direction of the force feedback deviation, the gripping pressure of the robotic gripper is increased or decreased, limiting it to a range from zero to the gripping pressure threshold after viscosity compensation. This generates optimized parameter adjustment instructions that comprehensively consider the effects of humidity, viscosity, and debris, enabling real-time compensation for abnormal contact conditions. These optimized parameter adjustment instructions may include, but are not limited to, angle adjustment instructions and gripping pressure adjustment instructions.

[0090] During the dynamic correction process of the robotic gripper based on the optimized parameter adjustment instructions, the stable benchmark values ​​of the gripper's angle and grasping pressure can be determined as the baseline closing angle and baseline grasping pressure. This set of baseline parameters can serve as reference actions and safety constraints for the deep reinforcement learning grasping policy network, enabling further optimization of grasping performance while meeting safety boundaries.

[0091] In some implementations, the target control parameters of the robotic gripper are determined based on the baseline closure angle, the baseline gripping pressure, and a pre-built deep reinforcement learning gripping policy network, including:

[0092] The state vector is input into a pre-constructed deep reinforcement learning grasping policy network, and the output is the closing angle reinforcement adjustment amount and the grasping pressure reinforcement adjustment amount. The state vector includes at least the humidity interference threshold, viscosity degree, force feedback contrast, debris embedding region features, baseline closing angle and baseline grasping pressure.

[0093] The target closure angle and the target gripping pressure are determined based on the closure angle enhancement adjustment amount, the gripping pressure enhancement adjustment amount, the baseline closure angle, and the baseline gripping pressure.

[0094] Specifically, a state vector is constructed using factors such as humidity interference threshold, viscosity level, force feedback bias, debris embedding characteristics, baseline closure angle, baseline grasping pressure, and optional contact area change rate and historical stability indicators. A pre-constructed deep reinforcement learning grasping policy network takes this state vector as input and outputs an action vector, which includes closure angle reinforcement adjustment and pressure reinforcement adjustment. The target closure angle and target grasping pressure are determined based on the closure angle reinforcement adjustment, grasping pressure reinforcement adjustment, baseline closure angle, and baseline grasping pressure.

[0095] In a specific example, the pre-built deep reinforcement learning grasping policy network can adopt an actor-critic structure, where: the actor network takes the state vector as input and outputs the closure angle reinforcement adjustment and the grasping pressure reinforcement adjustment; the critic network evaluates the grasp based on whether the grasp is successful, whether the garbage is broken or slipped, and stability metrics. Construct a reward function to update the policy gradient of the actor network.

[0096] In some embodiments, determining the target closure angle and the target gripping pressure based on the closure angle enhancement adjustment amount, the gripping pressure enhancement adjustment amount, the baseline closure angle, and the baseline gripping pressure includes:

[0097] Obtain the actual latency of the robot, and determine the dimensionless latency evaluation index based on the actual latency;

[0098] When the dimensionless time delay evaluation index is less than or equal to the preset time delay evaluation threshold, the closing angle enhancement adjustment amount is superimposed on the baseline closing angle, and the gripping pressure enhancement adjustment amount is superimposed on the baseline gripping pressure to obtain the target closing angle and the target gripping pressure.

[0099] If the dimensionless time delay evaluation index is greater than the preset time delay evaluation threshold, the baseline closure angle and the baseline grasping pressure are determined as the target closure angle and the target grasping pressure.

[0100] Specifically, the maximum allowable execution delay can be set according to the robot control cycle, and the allowable delay threshold for a single inference can be embedded into the control time window. The lightweight convolutional neural network and the deep reinforcement learning grasping policy network can be compressed and optimized to obtain a network structure that can achieve fast inference on the target hardware platform. The actual delay of multiple feature extractions and policy inferences can be statistically analyzed, and a dimensionless delay evaluation index can be calculated.

[0101] Specifically, set The actual delay of the i-th inference (unit: ms) Here, n represents the allowable latency threshold for a single inference attempt (in milliseconds), and n is the number of inference attempts counted. A latency evaluation metric is defined. Normalized average delay:

[0102] In the formula, the numerator and denominator have the same dimension, which is time. It is a dimensionless evaluation index.

[0103] When the dimensionless latency evaluation index is less than or equal to the preset latency evaluation threshold, the feature extraction and strategy reasoning process is deemed to meet the real-time requirements. The closure angle enhancement adjustment is then superimposed on the baseline closure angle, and the grasping pressure enhancement adjustment is superimposed on the baseline grasping pressure, yielding the target closure angle and target grasping pressure, which serve as the control parameters for this grasping operation. The target control parameters can be expressed as: ; .in, Indicates the target closure angle. Indicates the baseline closure angle. This indicates the adjustment amount for strengthening the closing angle. Indicates the pressure to capture the target. Indicates baseline gripping pressure. This indicates that the adjustment amount is strengthened by grasping the pressure. This indicates that the grabbing pressure will be limited to [0, The clipping function within the interval. This means that, based on the gripping pressure threshold after viscosity compensation, the adjustment result of the deep reinforcement learning output is ensured not to exceed the safe pressure upper limit defined by the viscosity mapping.

[0104] When the dimensionless time delay evaluation index exceeds the time delay evaluation threshold, in order to ensure real-time performance and security, the baseline closure angle and baseline grasping pressure are used as the control parameters for this grasping, and the corresponding status, action and result data are recorded in the experience.

[0105] In some embodiments, the method further includes:

[0106] During the grasping process of the robotic gripper, multi-dimensional grasping feedback data is acquired in real time. The multi-dimensional grasping feedback data includes at least pressure, contact area, and sliding trend.

[0107] A set of multi-frame crawling parameter vectors is constructed based on the multi-dimensional crawling feedback data;

[0108] The stability evaluation index of the crawling parameters is determined based on the set of multi-frame crawling parameter vectors;

[0109] If the stability evaluation index of the grasping parameters is greater than the preset stability threshold, the deep reinforcement learning grasping strategy network is given a low reward.

[0110] Specifically, under the conditions of meeting real-time constraints and adopting target control parameters, a timestamp is added to the final adjustment instruction generated in each inference, forming a time-stamped instruction set. An ordered execution queue is constructed using a priority sorting algorithm, and the sorted adjustment instructions are packaged and sent to the robotic gripper drive module to generate and execute the corresponding grasping action sequence, thereby realizing the grasping control of wet waste. During the robotic gripper's grasping action, multi-dimensional grasping feedback data is collected in real time from tactile sensors, and multi-dimensional grasping parameters related to pressure, contact area, and sliding trend are extracted to form a multi-frame grasping parameter vector set. The center vector C representing the ideal working state is obtained through clustering or statistical analysis methods, and the stability evaluation index of the grasping parameters is calculated. The mean squared deviation of each parameter vector relative to the center:

[0111]

[0112] in, It has the same dimensions as C (such as a vector composed of physical quantities such as pressure and angle). It is a Euclidean norm, therefore The dimension of is the square of the parameter dimension, which can be used as a stability indicator to measure the degree of fluctuation of the grasping parameters in the wet waste scenario.

[0113] Stability assessment indicators The stability threshold is the average squared deviation of the multi-frame capture parameter vector from the center vector of the ideal working state. It is determined based on a large number of stable capture samples under different wet waste scenarios. The distribution statistics are set to the upper bound of the distribution or a preset quantile. When When the current grasping parameters are less than a preset stability threshold, it is determined that the current grasping parameters have good dynamic adaptability to the wet waste scenario, and the smaller parameters will be adjusted accordingly. The successful capture result, along with the reward, is fed back as a high reward to the pre-built deep reinforcement learning capture policy network; when When the value exceeds the stability threshold or when grasping failures occur such as slippage or breakage, the corresponding sample is treated as a low-reward or negative-reward sample. The latest tactile features and their stickiness annotations are used as incremental samples. An incremental learning algorithm is used to update the support vector machine stickiness classification model online. At the same time, the pre-constructed deep reinforcement learning grasping strategy network is updated online or offline based on the state-action-reward data in the experience pool. This results in an improved stickiness judgment model and grasping strategy model, which are then used for subsequent tactile data processing and grasping control.

[0114] In a specific example, the pre-built deep reinforcement learning grasping policy network can adopt an actor-critic structure, where: the actor network takes the state vector as input and outputs the closure angle reinforcement adjustment and the grasping pressure reinforcement adjustment; the critic network evaluates the grasp based on whether the grasp is successful, whether the garbage is broken or slipped, and stability metrics. Construct a reward function to update the policy gradient of the actor network.

[0115] In this embodiment of the application, if the adaptability is insufficient after obtaining the dynamic adaptability judgment result, the system's adaptability in the wet waste scenario is improved by updating the classification boundary of the support vector machine and the deep reinforcement learning policy network parameters online.

[0116] Specifically, when When the tactile feature vector exceeds a preset stability threshold or the viscosity judgment accuracy decreases, the latest collected tactile feature vector and its labeled viscosity level are used as incremental samples. An incremental learning method is employed to update the support vector set and classification hyperplane parameters of the support vector machine, gradually shifting the viscosity classification boundary towards the new data distribution. The updated support vector machine is then used again for humidity / viscosity judgment and pressure threshold mapping, forming a front-end adaptive pathway of "tactile data - viscosity judgment - pressure control". Simultaneously, the state vectors within the corresponding time window, the actions output by deep reinforcement learning, stability indicators, and successful / failed grasping results are compiled into experience samples and stored in the experience pool. The system can update the parameters of the deep reinforcement learning policy network based on these experience samples during idle periods or through online fine-tuning, allowing its policy to gradually adapt to the latest wet waste distribution, viscosity characteristics, and force feedback features.

[0117] By combining the above-mentioned SVM incremental learning with deep reinforcement learning strategy updates, a closed loop is formed: tactile data → feature extraction and threshold mapping → deep reinforcement learning decision-making → real-time and stability evaluation → online model updates. This continuously improves the system's success rate, non-destructive nature, and long-term operational stability in complex wet and dripping garbage scenarios.

[0118] Please see Figure 2This is a schematic diagram of the structure of a robot grasping device based on tactile perception provided in an embodiment of this application. A second aspect of this application provides a robot grasping device based on tactile perception, the device comprising:

[0119] The acquisition module is used to acquire high-dimensional tactile data and perform noise reduction processing on the pressure distribution in the high-dimensional tactile data to obtain a smooth pressure signal.

[0120] The extraction module is used to extract features from the high-dimensional tactile data based on the smooth pressure signal to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features.

[0121] The first determining module is used to determine the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table, wherein the first mapping table represents the mapping relationship between humidity and mechanical gripper angle.

[0122] The second determining module is used to determine the stickiness of the waste based on the adjustment angle and the sticky stretching morphological characteristics, and to determine the gripping pressure threshold after stickiness compensation by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold.

[0123] The third determining module is used to determine the force feedback deviation based on the gripping pressure threshold after viscosity compensation, the characteristics of the debris embedding area, and the current gripping pressure.

[0124] The fourth determining module is used to determine the baseline closing angle and baseline gripping pressure of the mechanical gripper based on the force feedback deviation in response to the force feedback deviation being greater than a preset deviation threshold.

[0125] The fifth determining module is used to determine the target control parameters of the robotic gripper based on the baseline closing angle, the baseline gripping pressure, and the pre-constructed deep reinforcement learning gripping strategy network. The target control parameters include the target closing angle and the target gripping pressure.

[0126] The control module is used to control the mechanical claw to grasp wet waste according to the target control parameters.

[0127] The tactile perception-based robotic grasping device provided in the second aspect of this application can realize the various processes implemented in the above method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0128] Please see Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. A third aspect of this application provides an electronic device 3000, including a processor 3100 and a memory 3200. The memory 3200 stores machine-executable instructions that can be executed by the processor 3100. The processor 3100 can execute the machine-executable instructions to implement the above-mentioned robot grasping method based on tactile perception.

[0129] In some embodiments, this application also provides a machine-readable storage medium storing instructions that, when executed by a processor, cause the processor to implement the above-described tactile perception-based robot grasping method.

[0130] In some embodiments, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the tactile perception-based robot grasping method according to the above embodiments.

[0131] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0134] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0135] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0136] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0137] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0138] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A robot grasping method based on tactile perception, characterized in that, The method includes: High-dimensional tactile data is acquired, and the pressure distribution in the high-dimensional tactile data is denoised to obtain a smooth pressure signal. Based on the smoothed pressure signal, feature extraction is performed on the high-dimensional tactile data to obtain a target key feature set, including: inputting the smoothed pressure signal and its spatial neighborhood into a lightweight convolutional neural network to obtain an initial key feature set; combining a preset screening threshold and a support vector machine analysis method to screen the features in the initial key feature set, extracting regions with excessive humidity and local abrupt changes caused by debris embedding, and determining the target key feature set; the target key feature set includes humidity distribution features, viscous stretching morphology features, and debris embedding region features; Determining the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table includes: quantifying and evaluating the humidity distribution characteristics to obtain the degree of humidity interference; and determining the adjustment angle of the mechanical gripper based on the first mapping table when the degree of humidity interference is greater than a preset humidity interference threshold; the first mapping table represents the mapping relationship between humidity and the angle of the mechanical gripper. The stickiness of the waste is determined based on the adjustment angle and the viscous stretching morphology characteristics. The gripping pressure threshold after stickiness compensation is determined by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold. Determining the force feedback deviation based on the viscosity-compensated gripping pressure threshold, the characteristics of the debris embedding region, and the current gripping pressure includes: correcting the viscosity-compensated gripping pressure threshold based on the characteristics of the debris embedding region to obtain a debris-sensitive reference pressure; and determining the difference between the current gripping pressure and the debris-sensitive reference pressure as the force feedback deviation. In response to the force feedback deviation being greater than a preset deviation threshold, the baseline closing angle and baseline gripping pressure of the mechanical gripper are determined based on the force feedback deviation. Based on the baseline closure angle, the baseline gripping pressure, and a pre-built deep reinforcement learning gripping policy network, the target control parameters of the robotic gripper are determined, including: inputting a state vector into the pre-built deep reinforcement learning gripping policy network, and outputting a closure angle reinforcement adjustment and a gripping pressure reinforcement adjustment, wherein the state vector includes at least a humidity interference threshold, viscosity level, force feedback contrast, debris embedding region features, baseline closure angle, and baseline gripping pressure; determining the target closure angle and target gripping pressure based on the closure angle reinforcement adjustment, the gripping pressure reinforcement adjustment, the baseline closure angle, and the baseline gripping pressure; the target control parameters include the target closure angle and the target gripping pressure. The robotic gripper is controlled to grasp wet waste according to the target control parameters.

2. The method according to claim 1, characterized in that, The step of responding to the force feedback deviation being greater than a preset deviation threshold and determining the baseline closing angle and baseline gripping pressure of the robotic gripper based on the force feedback deviation includes: The optimal parameter adjustment instructions for the mechanical gripper are determined based on the magnitude and direction of the force feedback deviation. The optimal parameter adjustment instructions include angle adjustment instructions and gripping pressure adjustment instructions. During the process of the robotic gripper dynamically correcting itself according to the optimized parameter adjustment command, the stable reference values ​​of the robotic gripper's angle and gripping pressure are determined as the baseline closing angle and baseline gripping pressure.

3. The method according to claim 1, characterized in that, The step of determining the target closure angle and the target gripping pressure based on the closure angle enhancement adjustment amount, the gripping pressure enhancement adjustment amount, the baseline closure angle, and the baseline gripping pressure includes: Obtain the actual latency of the robot, and determine the dimensionless latency evaluation index based on the actual latency; When the dimensionless time delay evaluation index is less than or equal to the preset time delay evaluation threshold, the closing angle enhancement adjustment amount is superimposed on the baseline closing angle, and the gripping pressure enhancement adjustment amount is superimposed on the baseline gripping pressure to obtain the target closing angle and the target gripping pressure. If the dimensionless time delay evaluation index is greater than the preset time delay evaluation threshold, the baseline closure angle and the baseline grasping pressure are determined as the target closure angle and the target grasping pressure.

4. The method according to claim 1, characterized in that, The method further includes: During the grasping process of the robotic gripper, multi-dimensional grasping feedback data is acquired in real time. The multi-dimensional grasping feedback data includes at least pressure, contact area, and sliding trend. A set of multi-frame crawling parameter vectors is constructed based on the multi-dimensional crawling feedback data; The stability evaluation index of the crawling parameters is determined based on the set of multi-frame crawling parameter vectors; If the stability evaluation index of the grasping parameters is greater than the preset stability threshold, the deep reinforcement learning grasping strategy network is given a low reward.

5. A robotic grasping device based on tactile perception, characterized in that, The apparatus for applying the tactile perception-based robotic grasping method as described in any one of claims 1-4 includes: The acquisition module is used to acquire high-dimensional tactile data and perform noise reduction processing on the pressure distribution in the high-dimensional tactile data to obtain a smooth pressure signal. The extraction module is used to extract features from the high-dimensional tactile data based on the smooth pressure signal to obtain a target key feature set, which includes humidity distribution features, viscous stretching morphology features, and debris embedding region features. The first determining module is used to determine the adjustment angle of the mechanical gripper based on the humidity distribution characteristics and the first mapping table, wherein the first mapping table represents the mapping relationship between humidity and mechanical gripper angle. The second determining module is used to determine the stickiness of the waste based on the adjustment angle and the sticky stretching morphological characteristics, and to determine the gripping pressure threshold after stickiness compensation by combining the second mapping table. The second mapping table represents the mapping relationship between the stickiness and the gripping pressure threshold. The third determining module is used to determine the force feedback deviation based on the gripping pressure threshold after viscosity compensation, the characteristics of the debris embedding area, and the current gripping pressure. The fourth determining module is used to determine the baseline closing angle and baseline gripping pressure of the mechanical gripper based on the force feedback deviation in response to the force feedback deviation being greater than a preset deviation threshold. The fifth determining module is used to determine the target control parameters of the robotic gripper based on the baseline closing angle, the baseline gripping pressure, and the pre-constructed deep reinforcement learning gripping strategy network. The target control parameters include the target closing angle and the target gripping pressure. The control module is used to control the mechanical claw to grasp wet waste according to the target control parameters.

6. An electronic device, characterized in that, include: The memory is configured to store instructions; The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the tactile perception-based robotic grasping method according to any one of claims 1 to 4.

Images