Tactile Sensing for Soft Robotic Manipulators in 50 MPa Hydrostatic Pressure Environments

Deep‐sea exploration remains a challenging task as the extreme hydrostatic pressure environment, darkness, and suspended sediment launch severely hinder the capability of deep‐sea vehicles. As a complement to underwater camera, tactile perception becomes especially important in situations where machine vision is limited. However, tactile sensors utilized in deep sea, which should be able to detect pressure changes of only hundreds of pascals under high hydrostatic pressure, are still lacking. To tackle the challenge imposed by hydrostatic pressure, a simulated deep‐sea environment flexible sensor (SDEFS) is proposed, consisting of a force sensor array and a bending sensor based on hydrogels for tactile sensing in 50 MPa hydrostatic pressure environments. The force sensor is unaffected by the hydrostatic pressure and achieves high sensitivity of 82.62 N−1 under 100 MPa hydrostatic pressure. The SDEFS is utilized to classify objects based on the difference in hardness. It can accurately classify seven objects on the ground, and three objects in an underwater environment with hydrostatic pressure of 50 MPa, with total recognition accuracies of 98.3% and 96%, respectively. With high force measurement sensitivity and accurate recognition ability under water, the SDEFS is expected to provide very valuable haptic sensing and feedback in deep‐sea exploration.


Introduction
Deep sea (>200 m depth) encompasses over 95% of the world's oceanic volume and represents the largest and least explored habitat on Earth. [1]Marine research entered a period of rapid expansion since the 1960s, and countries around the world began large-scale investigation and research. [2][5][6] Deep-sea mining is endowed with more significant advantages than those of land-based mining.Most nodules lie on the seabed surface, which allows mining vessels to easily extract minerals and requires little or no seabed infrastructure. [7,8]Moreover, the deep-sea ecosystem is the largest biological habitat on Earth, and it has an extremely high biodiversity (gene diversity, species diversity, community diversity, and ecosystem diversity), which is a treasure house of new biological resources. [9,10]oreover, items in deep sea also provide a record of oceanic and climatic evolution spanning thousands of years, which plays an important role in the study of the ancient ocean and climate. [11,12]xploration has been hampered by the extreme conditions of the deep sea.Nearly 99% of deep-sea mineral resources are Deep-sea exploration remains a challenging task as the extreme hydrostatic pressure environment, darkness, and suspended sediment launch severely hinder the capability of deep-sea vehicles.As a complement to underwater camera, tactile perception becomes especially important in situations where machine vision is limited.However, tactile sensors utilized in deep sea, which should be able to detect pressure changes of only hundreds of pascals under high hydrostatic pressure, are still lacking.To tackle the challenge imposed by hydrostatic pressure, a simulated deep-sea environment flexible sensor (SDEFS) is proposed, consisting of a force sensor array and a bending sensor based on hydrogels for tactile sensing in 50 MPa hydrostatic pressure environments.The force sensor is unaffected by the hydrostatic pressure and achieves high sensitivity of 82.62 N À1 under 100 MPa hydrostatic pressure.The SDEFS is utilized to classify objects based on the difference in hardness.It can accurately classify seven objects on the ground, and three objects in an underwater environment with hydrostatic pressure of 50 MPa, with total recognition accuracies of 98.3% and 96%, respectively.With high force measurement sensitivity and accurate recognition ability under water, the SDEFS is expected to provide very valuable haptic sensing and feedback in deep-sea exploration.
unexploited, [13,14] and most pelagic fishes are poorly known due to the difficulty of capturing them. [15]There are still many knowledge gaps related to explorations in deep-sea environments, which require the aid of the submersibles of human occupied vehicles (HOVs), [16] remotely operated vehicles (ROVs), and [17] autonomous underwater vehicles (AUVs). [18]These submersibles, accomplished with suction samplers, rigid canisters, and industrial robotic manipulator arms, expanded the area of human operation. [19]Unfortunately, some seabeds are covered by sediments stemmed from bodies and waste of shallow organisms, which is an important source of nutrients for deep-sea organisms. [20]The sediments as well as dust are raised up by the submersible when it moves near the seabed.These foggy suspended particles will block the camera, rendering it impossible for the actuator to grasp objects accurately and distinguish the objects correctly.More seriously, it will interfere with the control of the submarine, resulting in device damage, hindering the submarine's ability of mining or sampling on the seabed.The sense of touch becomes particularly important in situations where vision is limited. [21]In addition, some of the minerals and aquatic organisms are damaged due to the devices being made of rigid components and have no haptic feedback capability, which are unsuitable for interactions with soft and brittle objects. [15]hus, equipping the robotic manipulator with soft tactile sensors is of great significance for deep-sea explorations.In the deep ocean, the hydrostatic pressure is as high as tens of megapascals, from which the sensors are required to detect pressure changes of only hundreds of pascals.[24] In addition, such high hydrostatic pressure will also cause water to enter the interior of the sensor, resulting in short-circuiting and incapacitation.Conventional rigid robots rely on pressure compensation methods to protect the sensors, [25] but these methods are unsuitable for flexible sensors.Therefore, the design and fabrication of stable tactile sensors that can distinguish pressure changes of hundreds of pascals under hydrostatic pressure over tens of megapascals is still challenging.
Fortunately, the ocean has provided inspiration regarding the design of such soft tactile sensors.For example, cephalopods are able to survive in the deep sea (up to 5000 m) without the protection of rigid shells because they maintain a balance between internal and external pressures due to their high-water-content muscles and skin. [26,27]Other flexible materials may have small bubbles during the preparation process, leading to deformation under huge hydrostatic pressure. [28,29]The hydrogel or other materials can exchange water molecules with the external environment, and these holes caused by bubbles will be filled by water, reducing material deformation.32][33][34] In this work, we design and fabricate biomimetic flexible tactile sensors to detect the contact force, bending angle, and identify the hardness of objects in the deep-sea environment.The simulated deep-sea environment flexible sensor (SDEFS) consists of a force sensor array and a bending sensor, and is covered by hydrogels, which are insensitive to hydrostatic pressure.The hydrogel with a homogeneous network and high crosslinking density is synthesized by a solvent-exchange strategy, which exhibits excellent antiswelling property.The modulus of the antiswelling hydrogel is further optimized by creating porous structures for improved measurement sensitivity of sensors.The force sensor array consists of Hall sensors on a flexible circuit board and an external hemispherical antiswelling hydrogel which is doped with NdFeB magnetic powder.Due to the excellent anti-swelling property and resistance to high hydrostatic pressure of the hydrogels, the force sensor is not affected by hydrostatic pressure and achieves a high sensitivity of 82.62 N À1 under pressure up to 100 MPa.The bending sensor is fabricated by flexible materials such as carbon nanotube (CNT), Ti 3 C 2 T x (MXene), and polydimethylsiloxane (PDMS).The bending sensor can work under hydrostatic pressure after calibration.The sensitivity of the bending sensor can reach 8.64°À 1 without hydrostatic pressure.With the increase of hydrostatic pressure, the sensitivities of the bending sensor will decrease, which are 5.45°À 1 and 2.26°À 1 under 50 and 100 MPa hydrostatic pressure.In addition, when the SDEFS is integrated into the soft gripper, the hardness of the grasped object can be distinguished by the outputs of the force and bending sensors.After trained by machine learning, the SDEFS can accurately classify seven materials on land, and three materials in 50 MPa underwater environment, with total recognition accuracies of 98.3% and 96%, respectively.With ultrahigh force measurement sensitivity and accurate recognition ability under water, the SDEFS is expected to provide very valuable haptic sensing and feedback in the fields of deep-sea exploration, resource exploitation, environmental protection, etc.

Design and Fabrication of SDEFS
Deep-sea environments are extremely challenging for both robots and creatures due to high hydrostatic pressure, low content of dissolved oxygen, and lack of light and nutrients.In spite, cephalopods are widely distributed in the sea, some of which survive at depths of up to 5000 m with soft bodies that can withstand high hydrostatic pressure and flexible brachial feet that can easily catch food with epidermal suction cups. [26]nspired by the cephalopods, we design an SDEFS that is composed of a 4 Â 3 array of force sensors and a bending sensor, mimicking the brachial foot nerve and the sucker tactile nerve, respectively, as shown in Figure 1a.The SDEFS provides the underwater vehicle with tactile and grasping information, which enables it to grasp soft and fragile objects in the deep-sea environment.SDEFS is designed as a sandwich structure, as illustrated in Figure 1b.A flexible printed circuit board (FPCB) along with a bending sensor are in the middle layer, covered with a PDMS silicone protective layer to prevent the seawater from soaking into it.The outer layer is coated with a layer of hydrogel, bonded to the silicone protective layer with adhesive, and a hollow hemisphere structure array is located on the hydrogel layer to serve as a force sensor (The detailed size and manufacturing processes of the SDEFS are available in Figure S1-S4, Supporting Information).Figure 1c shows the photograph of the SDEFS, and the soft gripper integrated with the SDEFS.The soft gripper with integrated the SDEFS is able to sense the contact force and hardness of the grasped object at 5000 m depth in deep-sea environments (50 MPa hydrostatic pressure).

Designs of Antiswelling Hydrogel and Bonding Adhesive
The most challenging issue for deep-sea soft tactile sensor is the extremely high hydrostatic pressure underwater.Because the hydrogel can store and retain a significant amount of fluid within their own structure, outer water can enter and fill in the pores and defects.As a result, the volume compression of the material by hydrostatic pressure can be reduced, rendering it a promising component for the fabrication of deep-sea soft tactile sensors.However, hydrogel suffers from swelling in the underwater environment, which will dramatically affect measurement sensitivity and accuracy of the sensor. [35,36]To tackle this problem, we synthesize an antiswelling hydrogel which can be used in deepsea environment.We use a solvent-exchange strategy to make an antiswelling hydrogel with citric acid (CA) as a crosslinking agent to enhance the strength of the hydrogel network.The hydroxyl groups of polyvinyl alcohol (PVA) and the carboxyl groups of CA were dissolved in dimethyl sulfoxide (DMSO) and crosslinked to form ester bonds during heating, forming the initial crosslinking.The prepared hydrogel shows a peak at 1740 cm À1 in Fourier-transform infrared spectroscopy (FTIR) spectra, represents the esterification between -OH group of PVA and -COOH group of CA, thus confirming successful crosslinking of the PVA hydrogel (Figure S5, Supporting Information).Moreover, due to the fact that DMSO is a strong hydrogen bond acceptor capable of forming hydrogen bonds with the hydroxyl groups of PVA, it affects the intra-and interchain hydrogen bonds among PVA segments by regulating polymer-polymer and polymer-solvent hydrogen bonds, thus affecting the conformation and arrangement of polymer chains.In the crosslinking step, the polymer network is switched to water, which is a relatively poor solvent and has a relatively weak hydrogen bond reception, as shown in Figure 2a.This step restores the intermolecular hydrogen bonds among PVA chains and forms a stiff and tough hydrogel. [37]Therefore, due to the high crosslinking densities derived from the homogeneous network and hydrogen bonds, the hydrogel has excellent antiswelling properties.
In order to further reduce the stiffness of the hydrogels, porous hydrogels (P-hydrogel) are fabricated by adding NaCl particles into the hydrogel precursor and then dissolving the particle in the process of solvent exchange.The NaCl particles are replaced by water, thus resulting in a porous structure.The moduli of P-hydrogels with different pore contents (i.e., NaCl weight fractions) are shown in Figure 2b.It is seen that the modulus decreases accordingly with the increasing NaCl weight fraction.For example, the modulus decreases from 0.31-0.12 to 0.05 MPa under NaCl contents of 0, 70, and 85 wt%.However, excessive pores will lead to a sharp decrease in the strength of the P-hydrogels after immersion in water, as shown in Figure S3, Supporting Information.Therefore, we choose the 70 wt% P-hydrogel for subsequent experiments.In addition, we mix NdFeB magnetic powder into the hydrogel to form a magnetic hydrogel (M-hydrogel) for tactile sensors.After sufficient stirring, the magnetic powder dispersed homogeneously in the hydrogel.The modulus and magnetic density flux of M-hydrogels with different mass ratios can be seen in Figure S5 and S6, Supporting Information.The influence of the hydrostatic pressure on the three kinds of hydrogels (original hydrogel, P-hydrogel, and M-hydrogel) is shown in Figure 2c,d changes of the three hydrogels are less than 1% when the external hydrostatic pressure increases from 0 to 100 MPa (Figure 2c).In the long-time immersion experiment, the effect of hydrostatic pressure on the hydrogels was also found to be negligible, which proves the ability of hydrogels to maintain structural stability in the deep-sea environment.In addition, the volume loss of porous hydrogels is greater under long-time immersion, which may be attributed to the contraction of the porous structure.It is noted that long-time immersion in seawater will result in the oxidation of the M-hydrogel.Furthermore, in order to obtain stable bonding between the outer hydrogel layer and the PDMS protective layer, an adhesive is used to bond these two layers together.Cyanoacrylates and 2,2,4-trimethylpentane are mixed in a certain weight ratio and then applied to the PDMS protective layer after sufficient mixing. [38]The mixture allows diffusion of the adhesive into the hydrogel and elastomers, leading to tough bonds while not forming rigid resin interlayers.The synergistic effects of physical entanglement due to adhesive interdiffusion and the ready formation of van der Waals and hydrogen bonds of   cyanoacrylates result in the high interfacial toughness of adhesive interfaces, [39] as illustrated in Figure 2e.The adhesive enables a reliable connection between the hydrogel and PDMS after stretching to 50% of their original length, as shown in Figure 2f.We also compare the influence of adhesive composition on the bond strength.As shown in Figure 2g, when the mixing fraction of cyanoacrylates is lower than 5 wt%, the adhesive cannot from reliable bonding.Whereas when the cyanoacrylates fraction exceeds 90 wt%, part of the cyanoacrylates will be cured by water in the air before combining with hydrogel.Moreover, excessive cyanoacrylates will affect the transparency of the bonding area.Therefore, a glue ratio of 10 wt% was selected for subsequent experiments.We investigated the effect of hydrostatic pressure on the bonding strength, as shown in Figure 2h.The bonded samples not immersed in water are named as initial state.Other samples are put into the pressure chamber with 0, 20, 50 MPa hydrostatic pressure applied.Samples using uncured hydrogels (without solvent exchange) are used for comparison to demonstrate the effect of water on bonding.It is seen that the bonding strength is nearly unaffected by the hydrostatic pressure, given that the shear strength only changes 2 kPa when the hydrostatic pressure varies from 0 to 50 MPa.Hydrostatic pressure has little effect on the bond strength.This is mainly because that hydrostatic pressure is a kind of pericentral extrusion pressure, which has no effect on the shear and tension of the bond layer.In addition, after immersing the samples into the pressure chamber with a pressure of 0, 20, 50 MPa for 72 h, their bonding strength barely changes, as shown in Figure 2i.This result indicates that the adhesive is an ideal adhesive for silicone and hydrogel in the deep-sea environment.

Hydrostatic Pressure-Balanced Simulated Deep-Sea Environment Force Sensors
Generally, traditional flexible sensors (including piezoresistive, piezoelectric, capacitive) are made up of hierarchical structures.The external pressure will be transformed into the change of the contact surface or volume of the sensing functional layer, resulting in change of output electrical signals.However, the hierarchical structure is not suitable for deep-sea environments, because the hydrostatic pressure will compress the functional layer to its limit.Even the flexible sensor with super wide detection range cannot withstand the deep-sea pressure of tens of megapascals.Instead, structures capable of balancing their internal and external hydrostatic pressure are less affected by hydrostatic pressure, allowing for the perception of external forces beneath the deep ocean.Therefore, we propose a novel simulated deep-sea environment force sensor which is almost impervious to hydrostatic pressure.The simulated deep-sea environment force sensor has a hemispherical array structure where the top of the hemisphere is fabricated by the M-hydrogel.The bottom of the hemisphere is fabricated by the P-hydrogel with a hole in the center.The hemispherical hydrogel is connected to a PDMS protective layer, which contains a FPCB in the middle, as shown in Figure 3a.When an external force F is applied to the hemisphere, the Hall sensors mounted on the printed circuit board will sense the deformation of the hemisphere according to the change in magnetic flux density B. The principle of force sensor measurement is as follows.
where M is the magnetization and μ 0 is the permeability of free space.The magnetic hydrogel is simplified as a cylinder, r is the radius of the cylinder, and h is the height of the cylinder.[42] Therefore, the variation of magnetic flux density and the applied force is approximately linear (the specific derivation process of the formula is shown in Supporting Information).We also simulated the performance of the simulated deep-sea environment force sensor with hollow structure (HS) and solid structure (SS) in the deep-sea environment, as shown in Figure 3b.The result demonstrates that the hydrostatic pressure of 100 MPa has negligible impact on the sensor, and it is still capable of sensing the external contact pressure in this extreme environment.Specific simulation settings are shown in Supporting Information.
The sensitivity of the simulated deep-sea environment force sensor S DFS can be represented as S DFS ¼ ðΔB=B 0 Þ=F, where ΔB and B 0 are the magnetic flux density induced by the external force and the initial magnetic flux density, respectively.Figure 3c shows the sensitivity of the simulated deep-sea environment force sensor with hollow structure and solid structure, indicating that the force sensor with HS has higher sensitivity than SS due to its lower structural stiffness.Specifically, the force sensor with HS has two different ranges of sensitivities.When the contact force F < 1 N, the sensitivity is 82.62 N À1 .During this period, the external force only affects the shape of the hollow hemispherical structure.When the contact force F > 1 N, the sensitivity is 20.48 N À1 .The decrease in sensitivity is mainly because the hollow hemispherical structure has been fully compressed, and the change in the magnetic flux density results from the compression of the base hydrogel layer.Generally, the crosstalk noise is affected by the distance between sensing units and the thickness of the base layer (as shown in Figure S8, Supporting Information); we selected distance of 6 mm and thickness of 2 mm for the sense array.As shown in Figure 3d, when an external force of 0.6 N is applied on one point of the force sensor array, the magnetic flux density change ΔB/B 0 of the contacted point is about 52.6%, while that of the adjacent point is only about 2.53%, indicating very low level of crosstalk in the sensor array.In addition, the response and recovery time are 30 ms.
Unlike the force sensors used in normal environments, deepsea environments pose more challenges, the most significant of which is hydrostatic pressure.The force sensor is tested under high hydrostatic pressure to explore its potential for deep-sea exploration.First, the influence of hydrostatic pressure on the baseline of the simulated deep-sea environment force sensor is presented.As shown in Figure 3e, the magnetic flux density of the sensor remains constant during the loading and unloading processes of hydrostatic pressure.Second, we demonstrate the capability of the force sensor to measure external contact forces in high-hydrostatic pressure environments.We also compared the effect of hydrostatic pressure on force sensors with the same structure but different materials, as shown in Figure S9, Supporting Information.We believe that the hydrogel material can exchange water molecules with the outside environment, and water fills the pores and defects in the original material, thus reducing the structural deformation under hydrostatic pressure.As for sensors made by PDMS, these internal pores are compacted by hydrostatic pressure, resulting in a reduction in the volume of the material, which affects the signal of the sensor. [43,44]In this experiment, underwater pressure loading is simulated by placing stainless steel blocks on top of the sensor array.As shown in Figure 3f, the sensor is still operating under highpressure conditions, and the sensor sensitivity is not affected by hydrostatic pressure.In addition, we measure the simulated deep-sea environment force sensor at 50 MPa hydrostatic pressure for 72 h.The change in magnetic flux density is negligible, as shown in Figure 3g.Finally, we composed a 4 Â 3 sensor array with 12 force sensors and tested its performance, as shown in Figure 3h (i), the signal output from the sensor under 50 MPa hydrostatic pressure.In the absence of external contact force, there is basically no change in magnetic flux density and the output signal of the sensor remains unchanged.In Figure 3h (ii), a stainless steel block with weight of 0.63 N was placed on the left side of the sensor array.As shown, the output signals of the nine sensing units in the left side increased.The contact force measured by the nine sensors is 0.597 N, and the increased magnetic flux density varies from 4.98% to 6.21%.Considering the slight buoyancy force, the measured contact force is in close agreement with actual value, indicating the measurement accuracy of the sensor array.The variation of magnetic flux density was due to the uneven position of the block and the impact of water flow (the effect of water flow on the sensor is shown in the Figure S10, Supporting Information).These results indicate that influence of hydrostatic pressure on the simulated deep-sea environment force sensor is negligible, and the simulated deep-sea environment force sensor still maintains high sensitivity under high hydrostatic pressure.

Design and Characterizations of Simulated Deep-Sea Environment Bending Sensors
Here, we design a strain sensor by leveraging the flexibility of CNT, MXene, and PDMS to measure the bending status of soft robotics in the simulated deep-sea environment, which is named as the simulated deep-sea environment bending sensor.MXenes are a family of 2D transition metal carbides or nitrides with excellent electrical conductivity and mechanical properties, [45] which are ideal materials for strain and bending sensors.[48] As shown in Figure 4a, the resistance variation of the bending sensor, which is induced by the change in the arrangement of conductive materials, can reflect the bending angle.When the bending sensor is exposed to high hydrostatic pressure, the space between the conductive materials decreases and new conductive paths are formed, which reduce the resistance of the sensor, as shown in (ii).When the sensor is bent under high hydrostatic pressure, the spacing of the carbon materials increases, and the original conductive paths are disconnected, resulting in greater resistance, as shown in (iii).Therefore, the bending sensor is calibrated at different underwater depths (hydrostatic pressure) and the bending angle can be calculated after calibration.The fabrication processes of the simulated deep-sea environment bending sensor are illustrated in Figure S2, Supporting Information.Briefly, a CNT suspension was mixed with MXene suspension by diverse mass ratios, and then the mixed suspension was filtrated to form a CNT/MXene composite film.Then the CNT/MXene composite films were immersed into the PDMS matrix.After PDMS polymerization, the membrane was peeled off and tailored into strip shape.The strip was then attached to the FPCB with conductive glue to form bending sensors.The morphologies of the bending sensor were characterized by a scanning electron microscope (SEM), as shown in Figure 4b.
The MXene is successfully combined with the CNT network.
Cracks are easy to form on the MXene film, while CNT network fills the gap of the MXene and reduces the occurrence of cracks.
The specific size of the sensor is shown in Figure S1, Supporting Information.We compared the effects of different volume ratios of CNT and MXene on sensor sensitivity.Figure 4c shows that increasing the content of MXene will improve the sensor sensitivity when no hydrostatic pressure is applied.The sensitivity of the simulated deep-sea environment bending sensor S DBS can be calculated as S DBS ¼ ðΔR=R 0 Þ=α, where ΔR is the resistance induced by bending, R 0 is the initial resistance of the sensor, and α is the bending angle of the sensor.The sensitivity is 0.07, 2.19, 8.64, 33.47, and 706.18°À 1 for the volume ratio of CNT/MXene of 0:1, 1:2, 2:1, 1:1, and 1:0, respectively.The bending sensor composed only of MXene has high sensitivity, but the measurement range is too narrow, while the sensitivity of the sensor consisting only of CNT is too low to be utilized in the curvature angle measurement.In addition, the influence of hydrostatic pressure on the initial resistance R 0 of the bending sensor is also an important factor to evaluate the sensor performance.We compared the variation of R 0 with different volume ratios of CNT and MXenes under 20 MPa hydrostatic pressure.As shown in Figure 4d, the value of R 0 decreases with increasing MXene content.The R 0 of MXene is reduced by 13.95% when under 20 MPa hydrostatic pressure, compared with 1.24% for the sensor composed only of CNTs.This phenomenon is due to the fact that MXene is a sheet material with a large gap between layers; even small external changes will lead to the connection and disconnection of the conductive paths between layers, rendering it highly sensitive and also subject to the influence of hydrostatic pressure.However, CNT is a fibrous material, and the gap between fibers is small, so the sensitivity is low, but it is less affected by hydrostatic pressure.Thus, considering the trade-off between the sensing sensitivity and the tolerance to hydrostatic pressure, the ratio of CNT:MXene = 2:1 was selected for the bending sensor.Figure 4e shows the influence of the loading and unloading process on the resistance variation of the simulated deep-sea environment bending sensor.The impact of the water flow in the tank during the pressurization process leads to the increase of the resistance of the sensor, which then decreases by 1.6% after stabilization.In addition, we also conducted a cycling test on the sensor, and after 1000 reciprocating bends, the resistance was only slightly drifted, as shown in Figure 4f.
The performance of the simulated deep-sea environment bending sensor under different hydrostatic pressures is shown in Figure 4g.The hydrostatic environment of 20 MPa has little influence on the simulated deep-sea environment bending sensor.A higher hydrostatic pressure will lead to the resistance drift of the sensor.The sensitivity of the bending sensor decays to 5.45°À 1 and 2.26°À 1 at 50 and 100 MPa, respectively.After calibration by depth, the bending sensor is capable of measuring the bending angle.However, the measurement sensitivity is rather low at 100 MPa hydrostatic pressure (Figure 4g).Thus, we use the sensor for bending angle measurement for underwater depth up to 5000 m (50 MPa hydrostatic pressure).

Object Recognition during Manipulation without Hydrostatic Pressure
We integrate SDEFS into a soft gripper to provide tactile and posture information during grasping.A hydraulic system is driven by a push rod motor to provide power for the soft gripper (the specific size is shown in Figure S11, Supporting Information).Figure 5a shows the soft gripper integrated with SDEFS and commercial sensors grasping Ecoflex, PDMS, and wood cubes.The soft gripper is mounted at the front end of a robot arm.The SDEFS is integrated in the right finger of the gripper, and commercial sensors including a pressure sensor and a bending sensor are stuck on the surface of the left finger.
The measurements of commercial sensors serve as reference for the SDEFS.Objects grasped by the gripper can be classified due to the difference in hardness.Under the same grasping condition of certain displacement of the actuator, the SDEFS has a different output when grasping objects with different hardness.It can be seen that the output value of the force sensor is lower and that of the bending sensor is higher when grasping soft materials such as foams compared with hard materials such as wood cubes.The reason for this phenomenon is that when grasping a soft object that deforms more easily, the gripper's curvature is larger and the deformation of the hemispherical structure on the force sensor is less.However, when grasping a hard object, it is difficult for the hard object to deform itself, which leads to less curvature of the gripper and a larger deformation of the hemispherical structure under pressure, as shown in Figure 5b.When grasping Ecoflex, PDMS, and wood cubes, the left finger is bent to an angle of 10.8°, 8.9°, and 7.1°, respectively.Figure 5c shows the real-time sensor data when grasping different objects.(i), (ii), (iii), and (iv) are the output signals from the simulated deep-sea environment force sensor, simulated deep-sea environment bending sensor, commercial bending sensor, and commercial pressure sensor, respectively.The data in the range of t 1 -t 2 , t 3 -t 4 , and t 5 -t 6 is the sensor output when the gripper grasps an Ecoflex cube, a PDMS cube, and a wood cube, respectively.The size of those cubes is the same (2 cm side length), and the average output of the simulated deep-sea environment force sensor is 0.45, 0.58, and 0.97 N for those three cases, while the average bending angle measured by the simulated deep-sea environment bending sensor is 10.4°, 9.16°, and 6.57°, respectively.It is noted that those data agree well with measurements from the commercial force sensor (0.51, 0.57, and 1.03 N) and the commercial bending sensor (10.13°, 9.47°, and 7.24°), validating the measurement accuracy of the deep-sea sensors.The data within t 7 -t 9 is the sensors output during the grasping-lifting up process of a fixed wood cube.From the moment t 8 , the robot arm is lifted upwards, as the wooden cube is fixed, resulting in slip between the soft gripper and the wooden cube.As shown from the data of deep-sea force and bending sensors, slip leads to significant signal fluctuations.We compare the output data of the SDEFS when grasping objects with seven different hardnesses under the same grasping condition, including foam (made of wood cellulose), Ecoflex, expandable polyethylene (EPE), PDMS, ethylene vinyl acetate copolymer (EVA), Rubber, and wood, the Shore hardnesses of these objects are 0.5, 1.3, 5.7, 24.3, 28.3, 32.6, and 96.2 HA, respectively (as shown in Figure 5d, S12, Supporting Information).SDEFS can distinguish the difference between soft and hard objects.The output of the force sensors increases as the hardness of the object increases.The simulated deep-sea environment force sensor outputs for foam, Ecoflex, EPE, and PDMS are 0.4, 0.45, 0.58, and 0.583 N, respectively.While that of rubber, EVA, and wood are 0.955, 0.964, and 1.053 N, respectively.The output of the bending sensors decreases as the hardness of the object increases (for instance 11.8°for foam and 6.57°f or wood).We further demonstrate the capability of SDEFS in recognizing the hardness of objects, after being trained by a convolutional neural network.The data of seven objects with different hardness, including foam, Ecoflex, EPE, rubber, EVA, PDMS, wood, are collected to train the convolutional neural network.The convolutional neural network is constructed by 24 convolutional layers and 12 fully connected layers.The specific network structure is shown in Figure S13, Supporting Information.The classification results of the data collected by the SDEFS are shown in Figure 5e.It can be seen that the SDEFS achieves an overall accuracy of 98.3%, and its accuracy toward classifying PDMS, rubber, EVA, and wood is as high as 98.8%.

SDEFS Utilized in Deep-Sea Exploration
The unique material and structure design of the SDEFS enable its high hydrostatic pressure tolerance and sensitive tactile sensing in deep-sea environments.As shown in Figure 6a, [49] a hydraulic system is driven by a push rod motor to provide power for the soft gripper.The entire experimental platform is fixed on an aluminum alloy support and placed in a high-pressure experimental cabin.It communicates with the outside world through a watertight joint.A camera is fixed in the cabin to provide image data during the experimental process.Detailed experimental methods are described in Figure S14, Supporting Information.
Figure 6b and S15, Supporting Information, show the measurement data of the sensor when grasping objects with different hardness under 50 MPa hydrostatic pressure (from left to right: foam, PDMS, and wood cube).Measurement results of the force sensor, the bending sensor, and the entirely sensor array are shown from top to bottom.Similar with ambient pressure condition, the grasping force is larger and the bending angle is smaller when grasping the hard object such as wood (blue lines), while for the soft object such as foam, the force signal is lower and the bending angle is larger (black lines).Figure 6c,d shows the outputs of the force sensor and the bending sensor, respectively, in the initial state and the state of grasping the wood block under 0, 50, and 100 MPa hydrostatic pressure.The force sensor is unaffected by the hydrostatic pressure.Under the hydrostatic pressure of 50 MPa, the output of the bending sensor decreases by 41.6% and 75% under the static and grasping states, while under the hydrostatic pressure of 100 MPa, the sensor decreases by 51.2% and 104.7%, respectively.It is noted that due to the impact of hydrostatic pressure at 100 MPa, the grasping operation did not lead to significant difference of R/R 0 (as shown in Figure S16, Supporting Information).Therefore, to ensure measurement sensitivity, we adopted the simulated deep-sea environment bending sensor for exploration at underwater depth up to 50 MPa.
We also conducted the hardness identification of the SDEFS in the high hydrostatic pressure environment, as illustrated in    fundamentally unaffected by hydrostatic pressure, the high recognition rate of the network can be ensured.

Conclusion
In summary, we have designed and fabricated an SDEFS to measure contact force, gripper bending angle, as well as classify objects based on their difference in hardness in the simulated deep-sea environments for underwater depth up to 5000 m.The SDEFS consists of a force sensor array, a bending sensor, and covered with antiswelling hydrogel.The hydrogel is an ideal substrate for deep-sea sensors due to its excellent antiswelling property and resistance to high hydrostatic pressure.The force sensor is unaffected by hydrostatic pressure and still maintains a high sensitivity of 82.62 N À1 at 100 MPa.The bending sensor can work under hydrostatic pressure after calibration.With the assistance of machine learning, the SDEFS can accurately recognize seven materials on land, and three materials in an underwater environment of 50 MPa, with total recognition accuracies of 98.3% and 96%, respectively.This work demonstrates the application potential of the SDEFS in the field of marine development and exploration.
Materials Synthesis: The solvent displacement method was adopted to prepare antiswelling hydrogel.Typically, 4 g PVA and 0.2 g CA with different mass were dissolved in 20 mL DMSO heated to 120 °C and stirred for 12 h.Then, the solutions were cooled to room temperature, and bubbles were removed by vacuuming.Finally, the transparent solutions (5 mL) were poured into acrylic molds and then immersed in 1 L of water at room temperature to form the hydrogel.The water was displaced by fresh water every 6 h to ensure the DMSO was thoroughly replaced.Unless otherwise stated, the hydrogel was prepared by solvent exchanging for 48 h.The P-hydrogel and M-hydrogel were prepared with the same process, but the difference is that 70%wt NaCl and 80 wt% NdFeB powder were added into the hydrogel precursor, respectively.
Simulated Deep-Sea Environment Force Sensor Fabrication: The hydrogel precursor was mixed with the NdFeB powder and poured it into the acrylic mold.The top of the hydrogel hemisphere was formed after solidification of the precursor and removed to another acrylic mold after solidification.Then, the NaCl powder was mixed into hydrogel precursor and poured into the second acrylic mold and immersed in water for 48 h to form a hemispherical hydrogel array, Finally, the hydrogel was magnetized and attached to the PDMS protective layer to form force sensors.
Simulated Deep-Sea Environment Bending Sensor Fabrication: A CNT suspension was mixed with Ti 3 C 2 T x suspension in diverse mass ratios and then the mixed suspension was vacuum filtrated.The filtrated film was heated to 60 °C to remove water and form a conductive film.Then, the PDMS was poured onto the conductive film and spun at a speed of 400 r min À1 for 1 min.Then, the CNTs/MXene/PDMS film was heated to 90 °C for 1 h.After PDMS polymerization, the filter membrane was peeled off and the composite film was cut into strips to form bending sensors.
Sensor Characterizations: A motorized Z-stage was used in combination with a force gauge (HP-5N by HANDPAI) to apply a well-defined force to the simulated deep-sea environment force sensor during the measurements.The force sensor data was collected from Hall sensors (MLX90393, Melexis) with an Arduino Uno via I 2 C protocol and sent to PC for further processing.The simulated deep-sea environment bending sensor was placed on 3D-printed arcs with the same length (30 mm) and different angles (0°, 30°, 60°, 90°, 120°, 150°), and the data were collected from a digital multimeter (Tektronix DMM6500) and sent to PC.For cyclic bending test, the simulated deep-sea environment bending sensor was placed on a precision moving platform (FlexTest Mini-S2-P, Hunan NanoUp Electronics Technology Co., Ltd) at a constant speed of 5 mm s À1 .
Manipulation Tests: A hydraulic system was driven by a push rod motor to provide power for the soft gripper, and the specific size is shown in Figure S11, Supporting Information.When grasping an object, the push rod motor was moved forward a fixed distance (10 mm, with a speed of 20 mm s À1 ) to maintain the same position in the soft gripper.When the object was released, the push rod motor was returned to the initial position (with a speed of 20 mm s À1 ) to ensure that the soft gripper was fully recovered.The soft gripper was mounted on a robot arm (UR5, UNIVERSAL ROBOTS) to perform grasping tasks.The SDEFS was integrated in the right finger of the gripper, and commercial sensors including a pressure sensor (Flexiforce sensor 25 lb, Flexiforce sensor) and a bending sensor (flex2.2,FLEX sensor) were attached on the surface of the left finger; the sensing performances of the two commercial sensors are shown in Figure S11, Supporting Information.
Simulated Underwater Tests: The soft gripper integrated with the SDEFS as well as the hydraulic system were sent into pressure chambers.Unless otherwise noted, experiments were performed at room temperature and 25 °C water temperature.The static pressure experiments were carried out in a small-pressure chamber (SHS500/1500-145S, 1 m in diameter and 2.5 m in height, up to 120 MPa hydrostatic pressure), the load and unload rates were 5 MPa min À1 , a watertight cable (8 cores, with a joint dimension of 5/8) was utilized for data transmission.Underwater experiments were carried out in a large pressure chamber (SHS1500/1000-150S, 2.5 m in diameter and 3 m in height, up to 120 MPa hydrostatic pressure), the load and unload rates were 5 MPa min À1 , a watertight cable (8 cores, with a joint dimension of 5/8) was utilized for data transmission, a watertight cable (8 cores, with a joint dimension of 7/16) for CAN communication of the motor provided power for the hydraulic system, and a watertight cable (8 cores, with a joint dimension of 7/16) for CCD data transmission, respectively.When the pressure chamber was pressurized, the soft gripper was bent due to the pressure difference.The hydraulic system was manually adjusted to ensure that the soft gripper was in the initial open state.

Figure 1 .
Figure 1.The deep-sea flexible sensor.a) The deep-sea sensor mimicking the structure of cephalopods.b) Diagram showing the structure of the sensor.c) Photograph of the sensor, (i) is the soft gripper integrated with the SDEFS, and (ii) is the photograph of the SDEFS.

Figure 2 .
Figure 2. Antiswelling hydrogel and underwater bonding adhesive.a) Schematic of antiswelling hydrogel preparation.b) Stress-strain curve and modulus of different porous hydrogels, the legend indicates the NaCl weight fractions.c) Influence of hydrostatic pressure on hydrogel volume.d) Long-time influence of hydrostatic pressure on hydrogel volume, the legend indicates the type of hydrogel and the hydrostatic pressure applied.e) Schematicshowing the adhesion between the hydrogel and PDMS.f ) Adhesive strength characterization of the hydrogel-PDMS bonding.g) Influence of mixing ratio of adhesive on bond strength.h) Effect of hydrostatic pressure on the bonding strength.i) Long-time influence of hydrostatic pressure on the bonding strength.The legend indicates that bonded samples not immersed in water are named as initial state.Other samples are put into the pressure chamber with 0, 20, 50 MPa hydrostatic pressure and uncured indicates samples using hydrogels without solvent exchange.

Figure 3 .
Figure 3. Design and characterizations of the simulated deep-sea environment force sensor.a) Structure of the force sensor.b) Simulation analysis of force sensor with SS and HS under 100 MPa hydrostatic pressure.c) Sensitivity of force sensors with SS and HS.d) Response time and recovery time of the sensor with 0.6 N contact force.e) Influence of hydrostatic pressure on the base value of the force sensor.f ) The capability of the force sensor to measure contact forces in hydrostatic pressure environments.g) Long-time influence of hydrostatic pressure on the force sensor.h) Working performances of the 4 Â 3 force sensor array, (i) is the signal output of the sensor when it was placed under 50 MPa hydrostatic pressure, (ii) a stainless steel block was placed on the left corner of the sensor array.

Figure 4 .
Figure 4. Design and characterizations of the simulated deep-sea environment bending sensor.a) Schematics of the simulated deep-sea environment bending sensor.b) SEM images of the simulated deep-sea environment bending sensor with different components.c) Effects of the mixing ratio of CNT and MXene on the sensitivity of the sensor.d) Changes of the sensor initial resistance with different volume ratio of CNT and MXene under 20 MPa hydrostatic pressure.e) Influence of the loading and unloading processes on the resistance variation of the sensor.f ) Cycling tests of the bending sensor.g) Sensing performances of the bending sensor under different hydrostatic pressure.

Figure 5 .
Figure 5. SDEFS utilized for grasping and distinguishing objects without hydrostatic pressure.a) A soft gripper integrated with SDEFS and commercial sensors grasps Ecoflex, PDMS, and wood cubes.The SDEFS is integrated in the right finger of the gripper, and commercial sensors including a pressure sensor and a bending sensor are attached on the surface of the left finger.b) Schematic illustration of the difference when the SDEFS grasps objects with different hardness.c) Real-time measurement data of grasping different objects, i, ii, iii, and iv show the schematics of the sensor and the measurement data from the simulated deep-sea environment force sensor, simulated deep-sea environment bending sensor, a commercial force sensor, and a commercial bending sensor, respectively.d) Sensor output when grasping objects with different hardness.e) The classification results of SDEFS.

Figure 6e .Figure 6 .
Figure 6e.The MPL Classifier neural network mentioned earlier is utilized to classify the hardness of grasping objects in the environment of 50 MPa hydrostatic pressure.The results are shown in Figure6f.Since the bending sensor will be affected by hydrostatic pressure, the recognition accuracy is 96%, while the recognition rate without bending sensor correction is 94%, (the complete confusion matrix is shown in FigureS17, Supporting Information).As the force sensor signal is