Bionic cross-medium grasping robot of rigid-flexible coupled variable stiffness material

Abstract

This invention relates to the field of cross-medium grasping technology, specifically to a biomimetic cross-medium grasping robot using a rigid-flexible coupled variable stiffness material. It employs a three-in-one system architecture of "material-structure-control," comprising: an end effector with a rigid-flexible coupled variable stiffness material layer and a biomimetic composite structure layer; a sensor array; an excitation device for applying a physical field to the rigid-flexible coupled variable stiffness material layer; and a controller. The controller is specifically configured to: fuse signals from the sensor array to identify the current medium type; predict future stiffness requirements based on historical sensor sequences; and control the excitation device to adjust the stiffness of the rigid-flexible coupled variable stiffness material layer, allowing it to switch between multiple modes, including a compliant mode, a rigid mode, and a locking mode. The robot achieves wide-range stiffness regulation through the rigid-flexible coupled variable stiffness material, stable adsorption in dry / wet environments through the suction cup-bristle composite microstructure, and adaptive medium control through multimodal perception and LSTM prediction.

Description

Technical Field

[0001] This invention relates to the field of cross-medium grasping technology, specifically to a biomimetic cross-medium grasping robot made of rigid-flexible coupled variable stiffness material. Background Technology

[0002] As the flexibility of automated pharmaceutical packaging production lines continues to increase, the operating environment faced by end effectors is becoming increasingly complex. Cross-media gripping, as the "last hurdle" restricting the level of automation in pharmaceutical packaging, still suffers from the following systemic defects:

[0003] (1) Material level: The response speed and stiffness variation range cannot be taken into account at the same time.

[0004] While heat-driven solutions (such as the Peking University lamprey suction cup and the Zhejiang University flexible gripper) can achieve a wide range of stiffness variations, their response speed is limited by the material's thermal conductivity and latent heat of phase change. The transition from "softening and bonding" to "hardening and locking" takes several seconds, far exceeding the critical time window for cross-medium gripping. Field-controlled solutions (such as magnetorheological fluids) offer fast responses, but their stiffness variation range is limited, and the excessive stiffness in the flexible state prevents them from fully conforming to microscopic morphologies. Existing technologies cannot simultaneously meet the requirements for both "fast" and "large" stiffness control within the same material system.

[0005] Structural level: Single biomimetic structures have an unavoidable media blind zone.

[0006] The gecko-inspired structure (micropillar array) exhibits significant adsorption attenuation underwater due to van der Waals forces shielding it from water molecules; the octopus-inspired structure (vacuum suction cup) has a high failure rate on dry, rough surfaces due to its inability to seal properly; while the hollow mushroom-shaped microstructure developed by Xi'an Jiaotong University has achieved preliminary success in both dry and wet environments, its high micro / nano fabrication cost, lack of intelligent control, and unresolved issues regarding bubble nucleation during dynamic processes remain. Existing biomimetic structures cannot simultaneously meet the comprehensive requirements of "dry / wet compatibility, low cost, and dynamic stability."

[0007] Control level: Lack of media perception and prediction capabilities.

[0008] Existing control systems are all open-loop or only possess propagation sensing capabilities, completely unaware of the surrounding environment. This results in most failures occurring within 0.5 seconds after a media switch. While the HMG system from Sun Yat-sen University / Huazhong University of Science and Technology integrates force sensing, it lacks predictive models such as LSTM, making it impossible to pre-adjust stiffness before a media switch. The paradigm shift from "post-event response" to "pre-event prediction" in control systems has not yet been achieved.

[0009] At the system level: the innovations are scattered and lack systematic integration.

[0010] Existing technologies have failed to integrate the advantages of shape locking, dry / wet compatibility, rapid switching, multi-mode integration, and zero energy consumption after locking into a systematic solution, resulting in the separation of "materials-structure-control" and the inability to generate synergistic effects.

[0011] Therefore, how to provide a systematic solution integrating "materials, structure, and control" to address the problem of "dispersed innovations and lack of system integration" in existing technologies has become an urgent issue to be solved. Summary of the Invention

[0012] The purpose of this invention is to provide a biomimetic cross-medium grasping robot made of a rigid-flexible coupled variable stiffness material, which aims to improve the stability of the contact interface during cross-medium grasping and achieve a balance between stiffness response speed and variation range.

[0013] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0014] A biomimetic cross-medium grasping robot made of a rigid-flexible coupled variable stiffness material includes:

[0015] The end effector has a rigid-flexible coupled variable stiffness material layer and a biomimetic composite structure layer;

[0016] A sensor array, integrated into the end effector, includes a conductivity sensor, a capacitance sensor, and a temperature sensor;

[0017] To drive the stiffness change of the rigid-flexible coupled variable stiffness material layer, the robot further includes an excitation device. The excitation device applies a physical field to the rigid-flexible coupled variable stiffness material layer, the physical field including but not limited to magnetic and thermal fields. In one specific embodiment, the excitation device includes a magnetic field generator (such as an electromagnetic coil) and an electric heating device (such as a flexible heating film), used to excite the magnetostrictive chaining effect of the magnetorheological fluid microcapsules and the thermally induced phase transition effect of the low-melting-point alloy microcapsules, respectively.

[0018] The controller is electrically connected to the sensor array and the end effector;

[0019] The rigid-flexible coupling variable stiffness material layer includes magnetorheological fluid microcapsules and low-melting-point alloy microcapsules, and a magnetorheological fluid enrichment layer is wrapped on the surface of the low-melting-point alloy microcapsules as a gradient transition layer, so that the coefficient of thermal expansion changes in three levels from the alloy to the silicone matrix. The thickness of the gradient transition layer is 0.5-2μm.

[0020] The biomimetic composite structure layer includes a suction cup array, a bristle array, and directional exhaust microchannels radiating from the center to the edge. The diameter of the suction cups gradually increases from the center to the edge, and the diameter, height, and spacing of the bristles gradually change from the center to the edge. The width of the directional exhaust microchannels is 5-10 μm and the depth is 2-5 μm.

[0021] The controller is configured to:

[0022] The current medium type is identified based on the signals from the sensor array;

[0023] Predicting future stiffness requirements based on long short-term memory neural networks;

[0024] Based on the prediction results, the rigid-flexible coupled variable stiffness material layer is controlled to switch between multiple modes, including compliant mode, rigid mode and locking mode;

[0025] Furthermore, when the conductivity sensor in the sensor array detects that the rate of change of the signal exceeds a preset threshold, the controller increases the clamping force within 50ms to cooperate with the directional exhaust microchannel to accelerate the expulsion of bubbles and avoid interface failure.

[0026] Preferably, the controller performs mode switching based on the predicted stiffness value: when the predicted value is 0.2-0.5MPa, it switches to compliant mode (initial contact, ensuring full fit); when the predicted value is 10-22.1MPa, it switches to rigid mode (high load gripping, resisting inertial impact); and when the predicted value is 5-10MPa, it switches to locking mode (long-term holding, zero energy consumption maintenance).

[0027] Preferably, the controller is further configured to immediately switch to rigid mode when the rate of change of the sensor signal exceeds 30% within 10ms, regardless of the prediction result of the long short-term memory neural network.

[0028] Preferably, the total number of parameters of the long short-term memory neural network is less than 1M, the model size is less than 4MB, and it is deployed on an embedded controller.

[0029] Preferably, the rigid-flexible coupled variable stiffness material layer is composed of the following components by mass percentage: 40%-50% silicone matrix, 8%-12% hydrogel reinforcing phase, 25%-35% magnetorheological fluid microcapsules, and 10%-20% low-melting-point alloy microcapsules. The magnetorheological fluid microcapsules can form a chain-like arrangement within milliseconds under the influence of a magnetic field, enabling rapid coarse adjustment of stiffness. The low-melting-point alloy microcapsules melt under electric heating, further enhancing stiffness within seconds, achieving a wide range of fine adjustment. The two components work in tandem in time and coexist in space, jointly achieving millisecond-level response and a stiffness variation range exceeding two orders of magnitude. In one specific embodiment of the present invention, a mixture of 45% silicone matrix, 10% hydrogel reinforcing phase, 30% magnetorheological fluid microcapsules, and 15% low-melting-point alloy microcapsules was used. Under test conditions (temperature 23±2℃, relative humidity 50±5%), the stiffness increased from 0.21MPa to 3.6MPa in about 55ms after a magnetic field was applied. After further applying electric heating for 1.2 seconds, the stiffness increased to 22.1MPa, with a total stiffness change of about 105 times.

[0030] Preferably, the biomimetic composite structure layer employs a suction cup-brisket composite microstructure, specifically including:

[0031] The suction cup array, used to provide macroscopic negative pressure, is the main source of adsorption. Its diameter gradually increases from 2 mm at the center to 5 mm outwards, and the lip thickness is 0.5 mm, showing a gradient distribution.

[0032] A bristle array, used to provide microscopic dry adhesion backup, has a diameter of 5-10 μm, a height of 10-20 μm, and a spacing of 10-20 μm, all varying in a gradient from the center to the edge. The bristle density in the central region is approximately 400 bristles / mm². 2 bristles disrupt the continuity of the water film at the microscopic scale, establishing direct "solid-to-solid" contact and providing a backup for dry adhesion. ;

[0033] The directional exhaust microchannels, with a width of 5-10 μm and a depth of 2-5 μm, are radially distributed from the center to the edge. The microchannels guide the rapid escape of bubbles at the gas-liquid interface, extending the bubble destruction time from 260 ms to more than 500 ms.

[0034] Preferably, the suction cup-brisket composite microstructure has a total adsorption force that satisfies the composite sealing model: ;

[0035] in, This represents the total normal adsorption force;

[0036] This represents the macroscopic negative pressure adsorption force, where... This indicates that the pressure difference between the inside and outside of the suction cup is generated by the vacuum system. Indicates the effective sealing area of ​​the suction cup;

[0037] This is the sum of the microscopic dry adhesion forces generated by all the bristles;

[0038] The interference coefficient is a mechanism parameter, ranging from 0 to 1, and was determined experimentally. In a specific embodiment of this invention, after gradient design and topography locking, testing was conducted. It can be reduced to around 0.12. It should be noted that this formula is mainly used for design analysis; real-time calculation is not required in actual control. ;

[0039] Physical meaning:

[0040] The two mechanisms work together perfectly without interference, and all the contributions of the bristles can be effectively added to the total adsorption force.

[0041] The two mechanisms completely interfere with each other, and the contribution of the bristles is completely suppressed (e.g., the deformation of the suction cup causes the bristles to be unable to make effective contact).

[0042] Actual value: Through experimental determination, this patent, through gradient design and morphology locking, will... The value decreased from 0.35 to 0.12, demonstrating that the mechanism interference was significantly reduced after optimization.

[0043] Preferably, the end effector is further provided with a quick-change interface, which adopts a conical positioning and steel ball locking structure, and the repeatability of positioning is ≤ ±0.03mm.

[0044] Preferably, the conductivity, capacitance, and temperature sensors in the sensor array include:

[0045] Conductivity sensors: 4, evenly distributed along a 90° circle, measuring range 0-50000μS / cm, accuracy ±2%;

[0046] Capacitive sensors: 2 units, arranged opposite each other, measuring range 1-100pF, accuracy ±1%;

[0047] Temperature sensor: 1 unit, center position, measurement range -20-80℃, accuracy ±0.5℃.

[0048] Preferably, the controller uses Kalman filtering to fuse the signals from the three sensors.

[0049] As a specific implementation method, the medium characteristics can be modeled as a state vector: ;in, It represents electrical conductivity (μS / cm). Represents the dielectric constant. The temperature (°C) is represented. Data from various sensors is fused using a sequential update strategy, and finally, Mahalanobis distance is used to match the data with a predefined medium feature library to output the medium type identification result. Those skilled in the art can set the observation matrix and noise covariance according to the sensor noise characteristics. Experiments show that this fusion method can achieve an accuracy of over 96% in identifying complex media such as pharmaceutical solutions.

[0050] Predicting the state of the medium at time k: The corresponding covariance is: ;in, Represents the state transition matrix (using a constant model I3×3). It represents the process noise covariance, reflecting minute fluctuations in the characteristics of the medium;

[0051] When new measurement value Upon arrival, calculate the Kalman gain: The status is then updated as follows: The covariance is updated as follows: ;in, Represents the observation matrix. Indicates the measurement noise covariance;

[0052] The system employs a sequential update strategy, performing an update each time a sensor data point is acquired: Conductivity sensor: Capacitive sensor: Temperature sensor: After three rounds of sequential updates, the optimal state estimate after fusion is obtained. ;

[0053] The matching degree between Mahalanobis distance and a predefined medium feature library is calculated: The media type with the smallest distance is selected as the identification result.

[0054] Preferably, the controller constructs a stiffness prediction model based on a long short-term memory neural network. It should be noted that those skilled in the art can also use other recurrent neural network structures with memory capabilities (such as gated recurrent units, GRUs) to achieve similar functions, but this application uses a long short-term memory neural network as a specific implementation method.

[0055] The long short-term memory neural network includes two LSTM layers and two fully connected layers, with a total number of parameters of less than 1M and a model size of less than 4MB.

[0056] The specific process is as follows:

[0057] Input processing: 50×3 timing input matrix Pass it into the first LSTM layer;

[0058] The first LSTM layer is calculated by iterating through 50 time steps, calculating each time step according to the LSTM unit formula, and outputting a shape of (50, 128).

[0059] The second LSTM layer is calculated by taking the output of the first layer as input, retaining only the hidden state of the last time step, and outputting a shape of (64,).

[0060] Fully connected layer computation: The vector passes through 256-dimensional and 64-dimensional fully connected layers sequentially, finally outputting a 10-dimensional vector. ;

[0061] Output analysis: The stiffness prediction values ​​corresponding to the next 10 time steps;

[0062] The prediction lead time is defined as the time difference between the predicted moment of stiffness change and the actual moment of medium switching. .

[0063] Preferably, a control method for a biomimetic cross-medium grasping robot using a rigid-flexible coupled variable stiffness material includes the following steps:

[0064] S1: Real-time acquisition of signals from a sensor array, which includes a conductivity sensor, a capacitance sensor, and a temperature sensor;

[0065] S2: Use Kalman filtering to fuse the signals and identify the current medium type;

[0066] S3: Based on a long short-term memory neural network, predict future stiffness requirements according to historical sensor sequences;

[0067] S4: Based on the prediction results, control the excitation device to adjust the stiffness of the rigid-flexible coupled variable stiffness material layer, so that it switches between multiple modes; and when the signal change rate of at least one sensor exceeds a preset threshold, increase the clamping force within a preset time, and immediately switch to the rigid mode regardless of the prediction results of the long short-term memory neural network.

[0068] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0069] This invention deeply integrates the material layer, structural layer, and control system, with the three mutually empowering each other to produce a synergistic effect of "1+1+1>3":

[0070] (1) Synergy between materials and structure: M-05 variable stiffness material is used to provide the V4.0 biomimetic structure with "morphology locking" capability, so that the biomimetic structure can fully fit the surface micromorphology in the softened state and its morphology is locked in the hardened state, and it can still maintain a certain adsorption force even if the negative pressure is completely removed.

[0071] (2) Synergy between structure and control: The directional exhaust microchannel of the biomimetic structure of the present invention V4.0 is not only a structural feature, but also linked with the control system, so that when the sensor detects a sudden change in conductivity, the control system can increase the clamping force 50ms in advance, and cooperate with the microchannel to accelerate the expulsion of bubbles.

[0072] (3) Synergy between control system and materials: The stiffness requirement output by the LSTM prediction model of this invention directly drives the composite excitation strategy of magnetorheological coil and alloy heating film, so that the control system can clearly determine when a fast response is needed (calling magnetorheological fluid), when a large-scale change is needed (triggering alloy phase transformation), and when zero maintenance lock is needed (structural reconstruction). Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0074] Fig. 1 This is a framework diagram of a biomimetic cross-medium grasping robot based on a rigid-flexible coupled variable stiffness material according to the present invention.

[0075] Fig. 2 This is a flowchart of a biomimetic cross-medium grasping robot control method based on a rigid-flexible coupled variable stiffness material according to the present invention. Detailed Implementation

[0076] 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.

[0077] Please see Figs. 1-2 The present invention provides the following technical solution:

[0078] Example 1: A biomimetic cross-medium grasping robot using a rigid-flexible coupled variable stiffness material, which adopts a three-in-one system architecture of "material-structure-control" to solve the technical problem of transient instability of the contact interface during cross-medium grasping.

[0079] (1) Material preparation and performance testing:

[0080] Weigh out 45g of silicone matrix, 10g of hydrogel reinforcing phase, 30g of magnetorheological fluid microcapsules, and 15g of low-melting-point alloy microcapsules according to the mass ratio, mix them evenly, and inject them into the mold. Cure them for 2 hours under a magnetic field strength of 0.3T and a temperature of 80℃ to obtain a rigid-flexible coupled variable stiffness material layer.

[0081] Test conditions: temperature 23±2℃, relative humidity 50±5%. After applying a magnetic field (0.3T), the material stiffness increased from 0.21MPa to 3.6MPa within 55ms; while maintaining the magnetic field, electric heating (2W power) was applied, and the stiffness increased to 22.1MPa after 1.2 seconds. After 1000 cycles of cyclic loading (5N load, 1Hz frequency), the stiffness decay rate was measured to be 6.2%. These data indicate that the material achieves millisecond-level response and a stiffness variation range exceeding 100 times, and exhibits good cyclic stability.

[0082] (2) A biomimetic cross-medium grasping robot made of a rigid-flexible coupled variable stiffness material, comprising:

[0083] The end effector has a rigid-flexible coupled variable stiffness material layer and a biomimetic composite structure layer;

[0084] A sensor array, integrated into the end effector, includes a conductivity sensor, a capacitance sensor, and a temperature sensor;

[0085] An excitation device is used to apply a physical field to the rigid-flexible coupled variable stiffness material layer, the physical field including but not limited to magnetic and thermal fields. In one specific embodiment, the excitation device includes a magnetic field generating device (such as an electromagnetic coil) and an electric heating device (such as a flexible heating film), used to excite the magnetorheological chaining effect of the magnetorheological fluid microcapsules and the thermal phase transition effect of the low-melting-point alloy microcapsules, respectively.

[0086] The controller is electrically connected to the sensor array and the end effector;

[0087] The rigid-flexible coupling variable stiffness material layer includes magnetorheological fluid microcapsules and low-melting-point alloy microcapsules, and a magnetorheological fluid enrichment layer is wrapped on the surface of the low-melting-point alloy microcapsules as a gradient transition layer, so that the coefficient of thermal expansion changes in three levels from the alloy to the silicone matrix. The thickness of the gradient transition layer is 0.5-2μm.

[0088] To verify the technical effectiveness of the gradient interface design, a control group experiment was set up: Group A used the gradient interface design of this invention (a 1μm magnetorheological fluid enrichment layer coated on the surface of low-melting-point alloy microcapsules), while Group B used conventional composite without a gradient interface (low-melting-point alloy microcapsules directly mixed with a silicone matrix). Under the same test conditions (temperature 23±2℃, relative humidity 50±5%, 5N load, 1Hz frequency), after 1000 cycles of loading, the stiffness attenuation rate of Group A was 6.2%, while that of Group B was 18.5%. The microcrack density at the interface (statistically determined by SEM images) was 67% lower in Group A than in Group B. These data indicate that the gradient interface design can reduce interface stress concentration by more than 50%, effectively suppressing interface debonding failure.

[0089] The biomimetic composite structure layer includes a suction cup array, a bristle array, and directional exhaust microchannels radiating from the center to the edge. The diameter of the suction cups gradually increases from the center to the edge, and the diameter, height, and spacing of the bristles gradually change from the center to the edge. The width of the directional exhaust microchannels is 5-10 μm and the depth is 2-5 μm.

[0090] The controller is configured to:

[0091] The current medium type is identified based on the signals from the sensor array;

[0092] Predicting future stiffness requirements based on long short-term memory neural networks;

[0093] Based on the prediction results, the rigid-flexible coupled variable stiffness material layer is controlled to switch between multiple modes, including compliant mode, rigid mode and locking mode;

[0094] Furthermore, when the conductivity sensor in the sensor array detects that the rate of change of the signal exceeds a preset threshold, the controller increases the clamping force within 50ms to cooperate with the directional exhaust microchannel to accelerate the expulsion of bubbles and avoid interface failure.

[0095] In this invention, the "50ms" fast response time is set based on a combination of the system sampling rate (typically 1kHz, corresponding to a sampling period of 1ms) and the control period (typically 1ms). Specifically, when the rate of change of the conductivity sensor signal exceeds a preset threshold, the controller can complete signal acquisition, threshold comparison, and clamping force command output within 1-2 control periods (i.e., 1-2ms), with the remaining time used for the mechanical response of the actuator (such as a solenoid valve or piezoelectric actuator). Experiments show that setting the fast response time to 50ms matches the average prediction lead of 48ms for the LSTM neural network, ensuring that clamping force adjustment is completed before bubbles form and begin to affect the interface seal (the critical time for bubble destruction is approximately 260ms). If the response time exceeds 100ms, the bubbles have fully developed, and the microchannel venting effect decreases significantly; if the response time is less than 10ms, the requirements for the actuator are too high, making it uneconomical in engineering. Therefore, 50ms is the optimal value determined after comprehensively considering sensor sampling, control calculation, actuator response, and the physical process of venting.

[0096] The preset threshold can be set according to the sensor noise characteristics and target detection sensitivity. In a specific embodiment of the present invention, a threshold is set as the signal change rate exceeding 30% within a 10ms window. This value is based on the following analysis: the conductivity change rate during normal medium switching is typically 5%-15% / 10ms, while the change rate during rapid bubble formation can reach 35%-50% / 10ms; setting the threshold to 30% can effectively capture dangerous events while avoiding false triggering. If the threshold is too low (e.g., 10%), it will lead to frequent false triggering; if the threshold is too high (e.g., 50%), the optimal intervention time may be missed.

[0097] Preferably, the controller performs mode switching based on the predicted stiffness value: when the predicted value is 0.2-0.5MPa, it switches to compliant mode (initial contact, ensuring full fit); when the predicted value is 10-22.1MPa, it switches to rigid mode (high load gripping, resisting inertial impact); and when the predicted value is 5-10MPa, it switches to locking mode (long-term holding, zero energy consumption maintenance).

[0098] The microscopic mechanism of the locking mode is as follows: The controller first controls the excitation device to ensure the material fully adheres to the microstructure of the target surface. Then, the magnetic field is removed and heating is stopped, causing the low-melting-point alloy microcapsules to solidify from a liquid state. The solidified alloy microcapsules form a spatial skeletal network within the silicone matrix, locking the matrix in the previously adhered morphological state. At this point, even if the external excitation (magnetic field, negative pressure) is completely removed, the material can still maintain close adhesion to the surface thanks to the internal alloy skeleton, thus achieving long-term retention in a zero-energy state. In short: the rigid mode relies on an external field to maintain shape, while the locking mode relies on the internal structure to memorize shape.

[0099] It should be particularly noted that the "rigid mode" and the "locking mode" of this invention are fundamentally different in terms of physical mechanism and energy consumption characteristics:

[0100] Rigid mode: The material is kept in a high-stiffness state by continuously applying an external physical field (magnetic field and / or thermal field). Once the external excitation is removed, the material will return to a low-stiffness state. This mode is suitable for scenarios that require rapid response and short-term withstand of large loads, but requires continuous power supply.

[0101] Locking Mode: Utilizing the solidification phase transition of low-melting-point alloy microcapsules, a permanent spatial skeletal network is formed within the material, "memorizing" and locking in the material's morphology. Even when external stimuli (magnetic fields, negative pressure) are completely removed, the material can still maintain a tight fit with the surface thanks to its internal alloy skeleton, achieving zero-energy long-term retention. This mode is suitable for scenarios such as static gripping and long-term fixation.

[0102] The switching between the two modes is automatically executed by the controller based on the stiffness requirements predicted by the LSTM: when the predicted value is 10-22.1 MPa, it switches to the rigid mode (high load gripping), and when the predicted value is 5-10 MPa, it switches to the locking mode (long-term holding). The two modes should not be confused, and the locking mode is one of the key innovations that distinguishes this application from existing thermally driven shape memory materials.

[0103] Preferably, for potential rapid failure scenarios (such as rapid bubble formation), the controller is further configured to immediately switch to rigid mode when the rate of change of the sensor signal exceeds 30% within 10 ms, regardless of the prediction result of the long short-term memory neural network. This threshold and window time can be adjusted according to the actual sensor sampling rate (e.g., 1 kHz) and control cycle (e.g., 1 ms), effectively avoiding grasping failures caused by the prediction model's failure to respond in time.

[0104] Preferably, the total number of parameters of the long short-term memory neural network is less than 1M, the model size is less than 4MB, and it is deployed on an embedded controller.

[0105] Preferably, the rigid-flexible coupled variable stiffness material layer is composed of the following components by mass percentage: 40%-50% silicone matrix, 8%-12% hydrogel reinforcing phase, 25%-35% magnetorheological fluid microcapsules, and 10%-20% low-melting-point alloy microcapsules. The magnetorheological fluid microcapsules can form a chain-like arrangement within milliseconds under the influence of a magnetic field, enabling rapid coarse adjustment of stiffness. The low-melting-point alloy microcapsules melt under electric heating, further enhancing stiffness within seconds, achieving a wide range of fine adjustment. The two components work in tandem in time and coexist in space, jointly achieving millisecond-level response and a stiffness variation range exceeding two orders of magnitude. In one specific embodiment of the present invention, a mixture of 45% silicone matrix, 10% hydrogel reinforcing phase, 30% magnetorheological fluid microcapsules, and 15% low-melting-point alloy microcapsules was used. Under test conditions (temperature 23±2℃, relative humidity 50±5%), the stiffness increased from 0.21MPa to 3.6MPa in about 55ms after a magnetic field was applied. After further applying electric heating for 1.2 seconds, the stiffness increased to 22.1MPa, with a total stiffness change of about 105 times.

[0106] Preferably, the biomimetic composite structure layer employs a suction cup-brisket composite microstructure, specifically including:

[0107] The suction cup array, used to provide macroscopic negative pressure, is the main source of adsorption. Its diameter gradually increases from 2 mm at the center to 5 mm outwards, and the lip thickness is 0.5 mm, showing a gradient distribution.

[0108] A bristle array, used to provide microscopic dry adhesion backup, has a diameter of 5-10 μm, a height of 10-20 μm, and a spacing of 10-20 μm, all varying in a gradient from the center to the edge. The bristle density in the central region is approximately 400 bristles / mm². 2 bristles disrupt the continuity of the water film at the microscopic scale, establishing direct "solid-to-solid" contact and providing a backup for dry adhesion. ;

[0109] The directional exhaust microchannels, with a width of 5-10 μm and a depth of 2-5 μm, are radially distributed from the center to the edge. The microchannels guide the rapid escape of bubbles at the gas-liquid interface, extending the bubble destruction time from 260 ms to more than 500 ms.

[0110] Preferably, the suction cup-brisket composite microstructure has a total adsorption force that satisfies the composite sealing model: ;

[0111] in, This represents the total normal adsorption force;

[0112] This represents the macroscopic negative pressure adsorption force, where... This indicates that the pressure difference between the inside and outside of the suction cup is generated by the vacuum system. Indicates the effective sealing area of ​​the suction cup;

[0113] This is the sum of the microscopic dry adhesion forces generated by all the bristles;

[0114] The interference coefficient is a mechanism parameter, ranging from 0 to 1, and was determined experimentally. In a specific embodiment of this invention, after gradient design and topography locking, testing was conducted. It can be reduced to around 0.12. It should be noted that this formula is mainly used for design analysis; real-time calculation is not required in actual control. ;

[0115] Physical meaning:

[0116] The two mechanisms work together perfectly without interference, and all the contributions of the bristles can be effectively added to the total adsorption force.

[0117] The two mechanisms completely interfere with each other, and the contribution of the bristles is completely suppressed (e.g., the deformation of the suction cup causes the bristles to be unable to make effective contact).

[0118] Actual value: Through experimental determination, this patent, through gradient design and morphology locking, will... The value decreased from 0.35 to 0.12, demonstrating that the mechanism interference was significantly reduced after optimization.

[0119] Preferably, the end effector is further provided with a quick-change interface, which adopts a conical positioning and steel ball locking structure, and the repeatability of positioning is ≤ ±0.03mm.

[0120] Preferably, the conductivity, capacitance, and temperature sensors in the sensor array include:

[0121] Conductivity sensors: 4, evenly distributed along a 90° circle, measuring range 0-50000μS / cm, accuracy ±2%;

[0122] Capacitive sensors: 2 units, arranged opposite each other, measuring range 1-100pF, accuracy ±1%;

[0123] Temperature sensor: 1 unit, center position, measurement range -20-80℃, accuracy ±0.5℃.

[0124] Preferably, the controller uses Kalman filtering to fuse the signals from the three sensors. As a specific implementation, the medium characteristics can be modeled as a state vector: ;in, It represents electrical conductivity (μS / cm). Represents the dielectric constant. The temperature (°C) is represented. Data from various sensors is fused using a sequential update strategy, and finally, Mahalanobis distance is used to match the data with a predefined medium feature library to output the medium type identification result. Those skilled in the art can set the observation matrix and noise covariance according to the sensor noise characteristics. Experiments show that this fusion method can achieve an accuracy of over 96% in identifying complex media such as pharmaceutical solutions.

[0125] Predicting the state of the medium at time k: The corresponding covariance is: ;in, Represents the state transition matrix (using a constant model I3×3). It represents the process noise covariance, reflecting minute fluctuations in the characteristics of the medium;

[0126] When new measurement value Upon arrival, calculate the Kalman gain: The status is then updated as follows: The covariance is updated as follows: ;in, Represents the observation matrix. Indicates the measurement noise covariance;

[0127] The system employs a sequential update strategy, performing an update each time a sensor data point is acquired: Conductivity sensor: Capacitive sensor: Temperature sensor: After three rounds of sequential updates, the optimal state estimate after fusion is obtained. ;

[0128] The matching degree between Mahalanobis distance and a predefined medium feature library is calculated: The media type with the smallest distance is selected as the identification result.

[0129] Experiments show that the recognition accuracy of a single sensor for drug solution A is only 82%, while the accuracy improves to 96% after fusion. Kalman filtering achieves superior recognition performance compared to a single sensor by dynamically adjusting the Kalman gain of each sensor and automatically increasing the contribution of the capacitive sensor when conductivity is ambiguous.

[0130] Preferably, the controller constructs a stiffness prediction model based on a Long Short-Term Memory (LSTM) neural network. In one specific embodiment, this LSTM network includes two LSTM layers and two fully connected layers. The input is three sensor signals (a 50×3 matrix) from the past 50 time steps (each time step is 1ms), and the output is the stiffness prediction value for the next 10 time steps. The network is trained using a mean squared error loss function and an Adam optimizer, and the training dataset is collected from 200 cross-media capture processes. Offline testing shows that the average prediction lead (i.e., the time difference between the predicted stiffness change and the actual media switching time) can reach 48ms. It should be noted that this prediction lead is used for offline evaluation of model performance and is not used as a real-time control parameter.

[0131] The prediction lead time is defined as the time difference between the predicted moment of stiffness change and the actual moment of medium switching. .

[0132] The model utilizes the timing pattern recognition capability of LSTM to capture upcoming handover events in advance from changes in sensor signals before the medium switch. The measured average prediction lead reaches 48ms, meeting the design goal of "50ms pre-tuning before handover".

[0133] (3) System-level performance verification: To verify the synergistic effect of the present invention, four sets of control experiments were set up (30 captures per set, with media randomly switched):

[0134] Group A (traditional baseline): ordinary silicone + planar structure + position control, success rate 16.7%;

[0135] Group B (Materials Only): The rigid-flexible coupling variable stiffness material (M-05) of this invention was used, but without a biomimetic structure. Open-loop switching was implemented, with a success rate of 63.3%.

[0136] Group C (Structure Only): The biomimetic composite structure (V4.0) of this invention was used, but without variable stiffness materials and adaptive control, with a success rate of 56.7%.

[0137] Group D (Complete Solution): Simultaneously employing the variable stiffness material, biomimetic structure, and LSTM predictive control of this invention, the success rate is 96.7%.

[0138] The above results show that the "material-structure-control" integrated solution of the present invention produces a significant synergistic effect, and its success rate is much higher than the simple summation of the independent effects of each group.

[0139] Preferably, as shown in Table 1, the present invention has significant advantages over the prior art in terms of materials, structure, control, and system.

[0140] Table 1. Comparison of data between existing technology and the present invention

[0141]

[0142] Example 2: A control method for a biomimetic cross-medium grasping robot using a rigid-flexible coupled variable stiffness material, comprising the following steps:

[0143] S1: Real-time acquisition of signals from a sensor array, which includes a conductivity sensor, a capacitance sensor, and a temperature sensor;

[0144] S2: Use Kalman filtering to fuse the signals and identify the current medium type;

[0145] S3: Based on a long short-term memory neural network, predict future stiffness requirements according to historical sensor sequences;

[0146] S4: Based on the prediction results, control the excitation device to adjust the stiffness of the rigid-flexible coupled variable stiffness material layer so that it switches between multiple modes; and when the signal change rate of at least one sensor exceeds a preset threshold, increase the clamping force within a preset time and perform rigid mode switching in priority over the prediction results of the long short-term memory neural network.

[0147] Example 3: LSTM Model Training and Validation:

[0148] Conductivity, capacitance, and temperature sensor data were collected during 200 cross-medium grasping processes (including media switching such as air→pure water, air→medicine, water→air, and medicine→water), with a sampling rate of 1kHz. For each grasping operation, the required stiffness values ​​for the next 10 time steps (10ms step size) were manually labeled.

[0149] The LSTM model is constructed as follows: First LSTM layer (50×3 input, 128 hidden units output, returns sequence); second LSTM layer (128 input, 64 output, only returns the last time step); fully connected layers are 256-dimensional and 64-dimensional; finally, a 10-dimensional vector is output. The mean squared error (MSE) loss function, Adam optimizer, learning rate 0.001, training for 200 epochs, and batch size 32 are used.

[0150] Validation set results: Medium identification accuracy is 94%, with an average prediction lead of 48ms (i.e., the predicted stiffness change time is 48ms ahead of the actual medium switching sensor signal change time on average). The model has approximately 0.9M total parameters and a size of approximately 3.6MB, and can be deployed on an ARM Cortex-M7 controller.

[0151] Example 4: Verification of Security Backup Strategy

[0152] In 50 random medium switching experiments, when using only LSTM prediction (without a backup strategy), four instances occurred where rapid bubble formation caused sudden changes in sensor signals (conductivity change rate exceeding 35% within 10ms). LSTM failed to trigger mode switching in time, resulting in capture failure. Adding a backup strategy: when the change rate of any sensor signal exceeds 30% within 10ms, immediately switch to rigid mode. In all four scenarios, the switch was completed before failure, improving the overall success rate from 92% (46 / 50) to 100% (50 / 50).

[0153] Example 5: Comparison of Experimental Data:

[0154] Comparison of gradient interface vs. non-gradient interface: Under the same cyclic loading conditions (1000 cycles, 5N, 1Hz), the stiffness attenuation rate of group A (gradient interface) was 6.2%, while that of group B (non-gradient interface) was 18.5%. The interface microcrack density of group A was 67% lower than that of group B, proving the stress reduction effect of the gradient interface design.

[0155] Comparison of microchannels vs. no microchannels: Under the same medium switching conditions (air → water), the bubble destruction time of group A (with microchannels) was 500ms, while that of group B (without microchannels) was 260ms, demonstrating the effectiveness of microchannel exhaust.

[0156] Clamping force linkage vs. no linkage comparison: Under the same bubble formation conditions (conductivity change rate exceeding 30% within 10ms), Group A (with linkage) achieved a 100% grasping success rate through clamping force combined with microchannel venting; Group B (without linkage) had a grasping failure rate of approximately 8% (4 failures out of 50 experiments), demonstrating the effectiveness of the structure-control linkage technology.

[0157] Example 6: Scheme Comparison:

[0158] Compared with the lamprey-inspired amphibious sucker (which uses an SMP array to achieve a 'rubber-glass' phase transition) published by Professor Yu Junzhi's team at Peking University in January 2026, this application proposes a different technical route—achieving millisecond-level response and a stiffness variation range of more than two orders of magnitude through a dual-mechanism time relay scheme of 'magnetorheological fluid microcapsules + low-melting-point alloy microcapsules'.

[0159] Example 7: Comparative verification of the synergistic effect of materials-structure-control:

[0160] To further verify the synergistic effect of the "material-structure-control" integrated architecture of this invention, the following comparative experiments were conducted. All experiments used the same end effector dimensions, test platform, and environmental conditions (temperature 23±2℃, relative humidity 50±5%). The grasping target was a pharmaceutical packaging bottle (material: glass, diameter 30mm, mass 50g), and the medium randomly switched between air, pure water, and liquid medicine (conductivity approximately 8000μS / cm), with each grasping involving at least one cross-medium switch. Each group underwent 30 grasping experiments (consistent with Group D conditions in Example 1).

[0161] Table 2. Comparison group design:

[0162]

[0163] Table 3 Experimental Results:

[0164]

[0165] 7.1 Synergistic Effect Analysis:

[0166] (1) Limited effect of individual innovation: The power of components B, C and D did not exceed 65%, indicating that no single dimension improvement can independently solve the complex problem of cross-media grasping.

[0167] (2) Pairwise combinations improve performance but bottlenecks still exist: The power of E, F, and G combined increases to 73%-80%, but there is still a 20%-27% failure rate. The specific failure mechanisms are as follows:

[0168] Group E (Materials + Structure, No Control): During rapid media switching (e.g., air → water), open-loop control cannot increase clamping force before bubble formation, resulting in the microchannel venting function not being activated and the interface seal being destroyed by bubbles. The bubble destruction time is approximately 260ms, while the open-loop response time is approximately 500ms.

[0169] Group F (materials + control, no structure): Although the control could predict and increase the clamping force in advance (within 50ms), the lack of directional exhaust microchannels prevented bubbles from escaping effectively, causing them to accumulate at the interface and weaken the adsorption force. Experiments showed that the bubble escape time increased from 500ms with microchannels to less than 260ms without microchannels.

[0170] Group G (Structure + Control, No Variable Stiffness): Control can guide exhaust, and the biomimetic structure can provide dry / wet adsorption, but the material without variable stiffness cannot achieve morphology locking on rough surfaces. When negative pressure fluctuates or fails temporarily, the material cannot maintain its conformal shape, resulting in a sharp drop in adsorption force.

[0171] (3) The complete solution of the three components achieves true synergy: the power of component H reaches 96.7%, significantly higher than any pairwise combination (maximum 80%). This improvement (16.7 percentage points) cannot be explained by the simple summation of the effects of each component, proving that there is functional coupling between the three components:

[0172] Material-structure coupling: The "locking mode" of variable stiffness materials enables the biomimetic structure to maintain its shape fit even after the negative pressure is removed, which is something that the structure alone cannot achieve.

[0173] Coupling of structure and control: The directional exhaust microchannel provides a physical pathway for the control system’s “increase clamping force every 50ms” command, and the two work together to effectively guide the bubbles out.

[0174] Coupling of control and materials: The stiffness requirement output by the LSTM prediction model directly determines when to invoke magnetorheological fluid (fast response) and when to trigger low-melting-point alloy (wide-range adjustment), realizing composite excitation of variable stiffness materials.

[0175] 7.2 Quantitative comparison with existing technologies:

[0176] The complete solution of this invention (Group H) is compared with cross-media grasping devices reported in the literature:

[0177] Table 4. Comparison of the present invention with existing cross-media gripping device technologies:

[0178]

[0179] The above comparison shows that, through systematic design, the success rate of the present invention in random switching scenarios across media is significantly better than that of the prior art, demonstrating outstanding substantial progress.

[0180] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A biomimetic cross-medium grasping robot made of a rigid-flexible coupled variable stiffness material, characterized in that, include: The end effector has a rigid-flexible coupled variable stiffness material layer and a biomimetic composite structure layer; A sensor array, integrated into the end effector, includes a conductivity sensor, a capacitance sensor, and a temperature sensor; Excitation device for applying a physical field to the rigid-flexible coupled variable stiffness material layer; And a controller, electrically connected to the sensor array, the excitation device and the end effector; The rigid-flexible coupling variable stiffness material layer includes magnetorheological fluid microcapsules and low-melting-point alloy microcapsules, and a magnetorheological fluid enrichment layer is wrapped on the surface of the low-melting-point alloy microcapsules as a gradient transition layer. The biomimetic composite structure layer includes a suction cup array, a bristle array, and directional exhaust microchannels radiating from the center to the edge. The controller is configured to: The signals from the sensor array are fused using Kalman filtering to identify the current medium type. Based on a long short-term memory neural network, future stiffness requirements are predicted according to historical sensor sequences. Based on the prediction results, the excitation device is controlled to adjust the stiffness of the rigid-flexible coupled variable stiffness material layer, so that it switches between multiple modes including compliant mode, rigid mode and locking mode. Furthermore, when the signal change rate of at least one sensor in the sensor array exceeds a preset threshold, the controller increases the clamping force within a preset time to cooperate with the directional exhaust microchannel to accelerate the expulsion of bubbles.

2. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The excitation device includes a magnetic field generator and an electric heating device.

3. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The thickness of the gradient transition layer is 0.5-2 μm.

4. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The directional exhaust microchannel has a width of 5-10 μm and a depth of 2-5 μm.

5. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The rigid-flexible coupling variable stiffness material layer is composed of the following components by mass percentage: 40%-50% silicone matrix, 8%-12% hydrogel reinforcing phase, 25%-35% magnetorheological fluid microcapsules, and 10%-20% low melting point alloy microcapsules.

6. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The input to the long short-term memory neural network is the sensor signals from the past 50 time steps, and the output is the stiffness prediction value for the next 10 time steps.

7. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The preset threshold is when the rate of change of the signal exceeds 30% within 10ms, and the preset time is 50ms; and when this condition is met, the controller immediately switches to rigid mode regardless of the prediction result of the long short-term memory neural network.

8. The biomimetic cross-medium grasping robot as described in claim 1, characterized in that, The diameter of the suction cup array gradually increases from the center to the edge, and the diameter, height, and spacing of the bristles gradually change from the center to the edge.

9. A control method for the biomimetic cross-medium grasping robot of claim 1, characterized in that, Includes the following steps: S1: Real-time acquisition of signals from a sensor array, which includes a conductivity sensor, a capacitance sensor, and a temperature sensor; S2: Use Kalman filtering to fuse the signals and identify the current medium type; S3: Based on a long short-term memory neural network, predict future stiffness requirements according to historical sensor sequences; S4: Based on the prediction results, control the excitation device to adjust the stiffness of the rigid-flexible coupled variable stiffness material layer, so that it switches between multiple modes; Furthermore, when the signal change rate of at least one sensor exceeds a preset threshold, the clamping force is increased within a preset time, and a rigid mode switch is performed prior to the prediction result of the long short-term memory neural network.

Images