Recent Advances on Underwater Soft Robots

The ocean environment has enormous uncertainty due to the influence of complex waves and undercurrents. The human beings are limited in their abilities to detect and utilize marine resources without powerful tools. Soft robots employ soft materials to simplify the complex mechanical structures in rigid robots and adapt their morphology to the environment, making them suitable for performing some challenging tasks in place of manual labor. Due to superior flexible and deformable bodies, underwater soft robots have played significant roles in numerous applications in recent decades. Meanwhile, various technical challenges still need to be tackled to ensure the reliability and practical performance of underwater soft robots in complicated ocean environment. Nowadays, some researchers have developed underwater soft robotic systems based on biomimetics and other disciplines, aiming at comprehensive exploration of ocean and appropriate utilization of unexploited resources. This review presents the recent advances of underwater soft robots in the aspects of intelligent soft materials, fabrication, actuation, locomotion patterns, power storage, sensing, control, and modeling; additionally, the existing challenges and perspectives are analyzed as well.


Introduction
The ocean covers about 70% of the Earth's surface area, much larger than the terrestrial world. [1]It contains not only various biological substances to be deeply studied but also rich natural resources such as oil and gas to be appropriately exploited.Adopting targeted treatments for deteriorating environmental pollution in the ocean is also one of today's global themes.However, the surface and bottom of the ocean are strongly influenced by currents, waves, and other impact elements.The underwater environment is extremely complicated and can easily lead to lifethreatening consequences.Therefore, humans need external tools to accomplish particular tasks (e.g., maintenance or monitoring applications, biological applications [2] ).Specifically, maintenance or monitoring applications include underwater pipeline inspections, offshore infrastructure maintenance, and condition monitoring.Biological applications include seabed and deep-sea exploration, sample collection from marine environments, and ecological aquatic phenomena monitoring and data collection. [3]To address this issue, it is important to design and manufacture easy-to-operate, high-efficiency, high-speed, and autonomous capabilities underwater robots.
Current autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs) are constructed from rigid materials, are motor-driven, and are traditional rigid underwater robots.By equipping AUVs and ROVs with underwater manipulators, which, in turn, are able to perform specific tasks (e.g., underwater pipeline inspection, sample collection, data monitoring, etc.), these underwater robotic manipulation systems are referred to as unmanned underwater vehicles (UUVs).Underwater rigid robots have stable structures and accurate transmissions with a certain number of degrees of freedom (DOF), and therefore can be programmed to precisely control the rotational or translational motion of each joint.However, due to the high modulus of elasticity of rigid materials, underwater rigid robots have a hard appearance and poor environmental adaptability, and are prone to harm marine organisms when they make contact with them.Traditional underwater jet propulsion methods can cause great disturbance to the underwater environment and influence marine organisms.In addition, traditional underwater rigid robots are large in size due to the presence of mechanical components such as control, transmission, and actuation, which makes it difficult to operate in narrow underwater spaces.
Soft robots are machines fabricated from compliant materials (polymers, elastomers, hydrogels, granules); they can operate with several different modes of actuation (e.g., pneumatic, electrical, chemical), and their motion can be either fast (>1 Hz) or slow (<0.1 Hz). [4]In addition, the bodies of soft robots are typically made of soft and=or expandable materials (e.g., silicone rubber) that can deform and absorb a large portion of the energy generated by a collision.Significantly, the soft materials used in robots can have a role in the actuation section, not simply for aesthetic, or protection.Therefore, if the main part of the robot is composed of a rigid structure, but soft materials are used for the propulsion part, it can be considered as a soft robot.However, if the robot uses a soft material for its protection, but the realization of the actuation is only related to rigid structures, it cannot be defined as a soft robot, for example, an UAV with a soft shell (e.g., Panasonic Ballooncam), an attachment imitating remora discs, [5] soft landing pads, [6] or soft propellers. [7]hese robots are rigid-flexible coupled UAVs using soft robot technology.
According to the application scenarios, soft robots can be categorized into underwater soft robots, terrestrial soft robots, and aerial soft robots.And as long as they are able to operate in in-water environments, they are considered underwater soft robots.Therefore, amphibious soft robots [8][9][10][11] are also considered as underwater soft robots.Soft robots with superior deformability, environmental adaptability, and simple drive control are well suited for application in underwater environments.Underwater soft robots can solve the performance defects of traditional rigid underwater robots, and have greater research prospects.More specific comparisons between underwater rigid robots and underwater soft robots will be expanded in Section 2.
Compared with air-contact terrestrial robots, underwater soft robots are floating in a water-wrapped environment and are always subject to buoyancy.Therefore, some empirical techniques applicable to terrestrial robots cannot be directly adapted to robots operating in the humid water environment.The development of underwater soft robots integrates fundamentals of materialogy, control science and engineering, computer science and technology, fluid mechanics, mechanical engineering, etc., aiming at exploring the state-of-the-art technologies in robotics.In recent years, as multidisciplinary cross-research is increasingly encouraged worldwide, the field of marine robotics has undergone continuous growth, and much more types of underwater soft robots are being investigated and developed for different purposes.Among them, underwater soft robots that mimic fish, [12][13][14][15] octopuses, [16,17] turtles, [18,19] or fallen leaves [20] based on the existing characteristics of organisms in nature perform relatively better.By combining with biomimicry, robots can make good use of biological compliance to accomplish tasks safely, efficiently, and accurately in complex marine environments.As there are many examples of biology outperforming machines in nature, nature is definitely the greatest teacher of functional design. [21]On the other hand, bioinspired soft robots provide an approach for subsea exploration in which the robots are composed of compliant materials that can better adapt to uncertain environments and take advantage of design elements that have been optimized in nature. [22]For example, researchers have mimicked the elastic and soft bodies of fish to develop soft robotic fish.Such robotic fish can move flexibly in dynamic water environment and can be used to accomplish monitoring, reconnaissance, salvage, and overhaul work in narrow or hazardous situations.
The earliest research in the field of soft robotics began in 1950; McKibben invented the fabricated pneumatic actuator. [23]uring the 1970 and 1990, robots based on other soft materials (e.g., granular materials, [24] elastomers, [25] fluids, [26] and gels [27] ) emerged.Since 1990s, materials science has even established the field of soft materials.The emergence of novel soft material technology has brought the capabilities of robots closer to those of natural organisms and has become one of the most rapidly developing research directions in the last decades. [28]Hence, the capabilities of underwater robotic systems to work in highly complex situations could be significantly improved, through well integration of its development with biomimetics, efficient control technology, and soft materials (with high elasticity and deformability).
Many scholars have reviewed the recent advances of soft robots, in aspects of materials, fabrication, actuation, modeling, sensing, control, etc. Subad et al. [21] presented a brief history of soft robots and recent innovations pertaining to tactile-sensor fusion for soft robots.Gariya et al. [29] reviewed the soft materials utilized for the manufacturing of soft robots.The actuators for soft robotics were successively reviewed by Boyraz et al., [30] El-Atab et al., [31] and Salunkhe et al. [32] More specifically, Walker et al. [33] reviewed the recent developments of pneumatic soft actuators, Gupta et al. [34] summarized dielectric elastomer (DE) actuators of soft robots, Chi et al. [35] summed up bistable and multistable actuators, and Marchese et al. [36] discussed the classification of three viable fluidic elastomer actuator (FEA) morphologies.Schegg et al. [37] and Rus et al. [38] comprehensively reviewed the simulation/design, modeling, and control of soft robotics.Sun et al. [39] provided a comprehensive overview of recent soft robotic locomotion research.Armanini et al. [40] reviewed the modeling of soft robots.Lee et al. [41] reported review of actuation, sensing, structure, control and electronics, materials, fabrication and system, and applications of soft robots.This description of soft robotics is intended to be a comprehensive technical summary of the field.Similarly, Pal et al. [42] discussed soft robotic control, sensing, and actuation.Subad et al. [21] presented a review of the evolutionary process of soft robotics technologies and systems.Whitesides et al. [43] suggested that soft robotics is more than a simple technical "tweak" of hard robotics; additionally, they promoted unique roles for chemistry and materials science in the "soft robotics" field.For underwater soft robots, the materials, fabrication, modeling, and control technologies are similar to those of terrestrial soft robots, whereas the actuation, locomotion pattern, power storage, and sensing technologies vary to some extent.As underwater soft robotics research is still in its infancy, relatively few reviews are available.Minaian et al. [44] reviewed the ionic polymer-metal composite (IPMC) artificial muscles in underwater environment, and provided an overview of the actuation, sensing, controls, and applications to soft robotics.Youssef et al. [2] discussed the current research progress and challenges in design, fabrication, modeling, and control of underwater soft robots.Hermes et al. [45] summarized the actuation and optimization for underwater soft robotics' crawling and swimming.All the above reviews provide inspirations and insights for this article's writing.
After a comprehensive search for relevant articles, we aim to focus on an all-inclusive and up-to-date review of underwater soft robots.Although the number of reviews on underwater soft robots is limited, there are also very informative reviews in recent years, which inspired us to write this article.Aracri et al. [3] discussed the technological and scientific gaps which represent the driving factors for the development of soft robotics.And they offered perspective on the development of sustainable soft systems, indicating the characteristics of the existing soft robots that promote underwater maneuverability, locomotion, and sampling.Two years ago, Fang et al. [46] recapitulated biomimetic inspirations, recent advances in soft matter materials, representative fabrication techniques, system integration, and exemplary functions for underwater soft robots.W.-S. Chu et al. [47] analyzed and compared the performance of SMA, IPMC, and lead zirconate titanate (PZT) actuators, and classified the swimming modes according to nonfish creatures.In order to distinguish the previous surveys, this article extends previous reviews by providing a more comprehensive summary of the latest technologies developed for underwater soft robots, which include intelligent soft materials, fabrication, actuation, locomotion patterns, power storage, sensing, control, and modeling; additionally, the existing challenges and perspectives are analyzed as well.Most particularly, we have also mentioned microswimming robots in Section 6.We hope that this review could be served as informative guidelines for experimentalists and practitioners engaged in the field of underwater soft robots.
According to the development trend, we organize the representative research studies according to chronological order.
In recent years, the development of underwater soft robots focused on materials and manufacturing, [18,[48][49][50][51][52][53] locomotion patterns, [36,[54][55][56][57][58][59][60][61][62][63][64][65][66] power storage, [15,55,60,67,68] sensing and control, [54,[69][70][71][72] and theoretical modeling. [73,74]Besides, communication technology for deep-sea robots is also a problem that remains to be solved.Therefore, in the following sections, state-of-the-art researches on underwater soft robots will be first categorized and analyzed in perspectives of intelligent soft materials, fabrication, actuation, locomotion patterns, power storage, sensing and control, and modeling.Second, the key communication technologies for underwater soft robots are also introduced in this review.A separate section regarding the recent advances on microswimming robots is also presented illustrating the size effect's impact on relevant technologies for underwater soft robots.Last but not the least, the outlook and challenges of underwater soft robots are prospected.The state-of-the-art status of underwater soft robots summarized in this review and derived inspirations from innovative works could contribute to the further development of advanced underwater robotic systems with better autonomy, intelligence, and practicality.

Underwater Soft Robots versus Underwater Rigid Robots
Underwater soft robots may encounter many unpredictable obstacles in underwater environment such as rocks, corals, a variety of fish, etc.In practical explorations, the underwater soft robot needs to be compatible with the natural environment and keeps itself safe during normal operations; hence, good environmental compatibility is an essential character for underwater soft robot.Unmanned submersible vehicles (ROVs and AUVs [75][76][77] ) can navigate for long distance for scientists to observe and study marine organisms.However, the propellers' rotation or the forward jetting of the submersible vehicles will lead to turbulence which may frighten the marine organisms. [78]In addition, conventional submersibles adopt mechanical structures and power devices; as a consequence, the employed hard and bulky shells make it difficult for submersibles to be integrated into the marine environment and dive into relatively complex terrain, not conducive for clear observation of marine organisms as well.What is more, some certain submersibles need to be tethered, which has limited their operational flexibility as well. [79]onventional underwater rigid robot is developing and maturing during the last decades; crucial parameters such as diving depth, moving speed, and working performance (e.g., endurance, weight, portability, etc.) have improved a lot.Also, the corresponding commercialized applications have expended to areas including resource exploration, [80,81] safety inspection, [82] and search and rescue. [83]In contrast, most underwater soft robots are still at the stage of laboratory research and development, and will take some time to achieve comprehensive commercialization.However, underwater soft robots are expected to well compensate for the limitations of conventional underwater robots in certain practical applications in case they can overcome some technical difficulties (e.g., poor battery life, poor pressure resistance, poor controllability, slow speed, limited communication, and complex fabrication) that exist in current research.
The comparison of some underwater rigid robots and underwater soft robots are listed in Table 1.We measured the effectiveness of the robot locomotion by using the cost of transport (COT) which represents the amount of energy that is needed for traversing from one place to another. [84]COT is a dimensionless metric that allows locomotor efficiency comparisons across different robots and organisms.In addition, all COTs in Table 1 refer to TCOTs (total cost of transport), [85] which are given by COT = P/mgv, [22,84,86,87] where P (W) is the instantaneous power consumption, m (kg) is the robot's mass, g = 9.81 N kg À1 is the gravitational acceleration, and v (m s À1 ) is the speed in the forward direction.Calculating COT was not possible when any of mass, velocity, or power of robot was unreported.In some cases, soft materials have a natural advantage in simulating continuous compliant kinematics and can also periodically store and release mechanical energy during deformation, which enables soft robots to achieve low COT and high durability. [88]Additionally, Figure 1 visualizes the speed and size of various underwater rigid robots and underwater soft robots in an intuitive view.Obviously, underwater soft robots have a wider speed and size span than traditional underwater rigid robots.Soft robots which explore at deep sea (with depths up to 11 000 meters) have to overcome a series of challenges such as ultrahigh external pressure, extremely low light, and remote communication. [15]Besides, some underwater soft robots are designed to be equipped with sensors that can withstand high pressure and vision sensing modules for capturing real-time images in deep-sea low-light environments. [21]In scenario of exploiting marine resources, the soft robots are implemented to accomplish specific actions like grasping, salvaging, drilling, etc.While in task of marine environment protection, the robots are expected to be compatible and environmental friendly with the surroundings, [3] ensuring the working efficiently and sustainability in the high-salinity and highmobility ocean.

Existing Challenges
The last decades have witnessed the mature development of terrestrial robots; Boston Dynamics is a signature representative.The company has been dedicated on the developments of humanoid bipedal machines, [93] dog-like quadrupedal robots, [94,95] and Handle, a two-wheeled robot for cargo handling. [93,96]These robots can not only highly reproduce complex gaits (e.g., running, jumping, backflips, etc.), but also react to the sudden changes in surrounding environment intelligently, making them widely applied in the real world.
Compared with conventional terrestrial robots, the research on underwater robots over the last decade evolved relatively slower, mainly attributed to the uncertainty and complexity of underwater environment.Challenges like ocean currents and low visibility make it difficult for underwater robots to have the same situational awareness as walking on the ground or flying in the air.Besides, traditional gait control and mechanical design methods for terrestrial robots could not be applied on underwater robots directly; to achieve autonomous navigation, modified actuation and motion control strategy are required.Particularly, high-performance waterproof and antipressure treatments are prerequisite for the electronic components in deep-sea underwater robots.In addition, thruster-based underwater robots must proactively and continuously control their thrusters and buoyancy to maintain a constant posture in the water, which leads to a challenge of high energy consumption.
Furthermore, due to the inherent characteristics of ocean and environmental interference factors like surges, traditional communication methods (e.g., underwater electromagnetic wave communication, underwater hydroacoustic communication) will be seriously affected by environmental noise [97] and highly attenuated.The submerged quantum communication is one emerging technology to tackle above problems; a detailed performance comparison between two traditional wireless communication methods and the emerging submerged quantum communication is presented in Table 2.
The development of underwater soft robots can be traced back to the year of 1950 when artificial muscles were first invented; [23] since then, it has become the frontier research hotspot along with numerous challenges summarized as follows: 1) Soft materials have become an influential factor as well as the challenge in developing high-performance underwater soft robots.Novel bionic soft materials / structures have raised new requirements on advanced manufacturing technologies like 3D/4D-printing.
2) Soft robots are continuum robots and their infinite DOFs increase the complexity of kinematic modeling and control.Hence, more accurate analytical models and effective simulation methods need to be investigated.3) A large variation exists between the simulation and practical experiment results, mainly due to the influence of complex flow disturbance in underwater environment.New control algorithms suitable for complex dynamic underwater environments are expected to be developed.4) Sensing technologies (e.g., acoustic, optics, electromagnetic, and bionic) have certain drawbacks when adopted in underwater environment; most conventional sensors cannot be applied in underwater environment directly; also flexible sensors that are compliant with underwater soft robots are also highly demanded.5) Soft actuators made of intelligent soft materials are superior in some respects, but still less efficient than conventional actuators  [343,344]  and have a limited range of applications.6) Effective communication for untethered underwater soft robot is very important for target detection, localization, and identification.However, traditional aquatic communication technologies cannot be directly adapted to underwater applications due to the change of the medium.7) The lack of light, huge water pressure, inadequate power supply, poor communication, and limited vision during the diving into deep sea have limited the real working depth of underwater soft robots.
For the existing challenges mentioned above, the sensing technology and underwater communication technology are technical barriers that have more largely limited the performance of the underwater soft robot, restricting it from becoming more intelligent, in comparison with the problems in structural design and environmental coupling.Therefore, a more detailed description of these two challenges is expanded further.
The underwater rigid robot is fabricated of rigid material and does not have large deformation capability, so the sensors usually carried by underwater rigid robots are also rigid.However, soft robots are made of soft materials, [98,99] and to be better compatible with the dynamic deformation of soft robots, flexible sensors with elastic deformation of high coconformal capability are more suitable to be carried by underwater soft robots. [100]Therefore, the sensing technology of rigid underwater robots cannot be directly applied to underwater soft robots.However, the development of high-performance flexible sensors suitable for underwater soft robots is still facing significant challenges; as the flexible sensors are made of soft materials, there is a hysteresis effect; [101,102] to overcome the negative impacts brought by the hysteresis effect, it can be achieved through structural optimization, [103] material selection, [104] and advanced control methods. [105]nderwater communication methods for traditional rigid robots (e.g., AUVs) mainly include cable-based communication, [106] electromagnetic wave communication, [107] underwater optical communication, [108] and hydroacoustic communication. [109]Underwater cable communication is the most mature technology for underwater communication at present, and it is mainly applied in the normal underwater communication of ROV and cable (or microfiber optic cable) submersibles.This method has the advantages of large communication capacity, resistance to electromagnetic interference, and good confidentiality.The realization of these functions requires greater tensile strength, bending displacement, corrosion, and aging resistance for cables and fiber optic cables.However, due to the limitation of cable/optical cable length, it is impossible to carry out longer distance underwater communication, and the submarine's underwater operation is seriously limited, and all the operations need to be coordinated by the mother ship.Both electromagnetic wave communication and optical communication have limited underwater communication ranges.The former is severely affected by strong attenuation, which results in smaller propagation distances, while the latter depends on water turbidity.Hydroacoustic communication is now the dominant technology for underwater communication because it can achieve much longer communication ranges.Ideally, traditional AUV underwater communication methods could be applied to underwater soft robots.However, underwater soft robotic is now tending to be miniaturized, lightweight, and highly flexible, [12,110] so the load capacity of underwater soft robots is limited, [15,111] and the traditional bulky communication equipment cannot be directly mounted on the underwater soft robots.Without affecting the dynamics and kinematics performance of the soft robot, the communication equipment of the underwater soft robot should satisfy the requirements of long transmission distance as well as the requirements for designing the communication equipment with small size and lightweight.
In the last decade, underwater soft robots have been developed and have different challenges to overcome; in this article, we will first review selected significant achievements chronologically.In the years 2012-2017, underwater soft robotics technologies have been rapidly developed, with numerous relevant technical achievements reported, and a series of functional elastic materials, actuation, sensing, and fabrication technologies are being invented (Figure 3).
Kim et al. [18] developed a turtle-inspired biomimetic smart soft composite (SSC) structure actuator, referring to the structure of a sea turtle flipper and its lift-based swimming patterns (Figure 3a).The actuator has smoother, softer locomotion with more DOF than conventional ones.The turtle-inspired soft robot equipped with invented actuator is capable of reaching a maximum speed of 11.5 mm s À1 .This research is also an innovative application of SMA with superior performance.
Onal et al. [28] designed a bioinspired soft robotic snake that contains four bidirectional FEAs, and two miniature lithium polymer batteries for electrical energy (Figure 3b).It also owns computational components and well controllability, taking advantage of fluid dynamics to achieve snake-gait nearcontinuum morphological locomotion.The snake-inspired soft robot can attain an average locomotion speed of 19 mm s À1 .The authors innovatively harnessed the slow dynamics of soft actuators in order to convert electrical square wave input into mechanical sinusoidal output.
Marchese et al. [12] designed a soft robotic fish capable of independent locomotion which also utilized fluid elastic actuators (FEAs) (Figure 3c).The robot is also capable of mimicking the swimming and escape maneuvers of fish when coupled with self-designed control algorithms.In this research, they abandoned simple rectangular shape and created FEAs that conform to the complex anatomical shape of the fish.
Urai et al. [115] proposed an underwater soft robot inspired by the morphological features of rays (Figure 3d).They mimicked both their radially skeletal structure (with independent actuators) in the design of each bone and the compliance of their fins.The developed robot was shown to be capable of ray-like flexible locomotion adopting a simple open-loop control framework.
Cianchetti et al. [17] developed the soft eight-arm OCTOPUS robot (Figure 3e).Soft robotic octopus arm mimicked the adaptability of octopus' morphology, soft and compliant skin with longitudinal and transverse muscular structures, achieving variable stiffness and good dexterity.The robot can automatically react to the environment through extensive computation without the need of complex manual commands (to control the eight claws).It achieved breakthrough by combining new technologies into a unique platform.Unlike the four abovementioned soft robots that carry their LiPo battery power, the frontal arms of this octopus-inspired robot are equipped with motor-driven cable actuators and SMA actuators which need to be powered by external cables.
Li et al. [55] designed a cableless soft-bodied robotic fish using DE membrane as muscle, silicone film as fin, and silicone frame as elastomer to mimic the structural and locomotion mechanism of a manta ray flapping its pectoral fins (Figure 3f ).The robotic fish is powered entirely by a soft electroactive structure made of DE and ionic conductive hydrogel.Moreover, it can swim at the speed of 6.4 cm s À1 (0.69 BL s À1 , body length per second), which is much faster than previously reported cableless soft robotic fish driven by soft responsive materials.This research demonstrates that the high voltage requirement for actuating the DE is compatible with the aqueous operating environment, and also demonstrates DE's superior performance in the field of underwater soft robots.
Till 2017, the technological base of soft material actuators, sensors, and control algorithms has been developed to a higher stage.Hence, researchers are no longer limited in developing soft robots that just mimicking single movement of creatures in nature, but trying to develop robotic systems that can achieve different locomotion patterns. [19,117,122,123]The gait of robots evolved toward more flexibility, thus owning features alike their original creatures.In addition, they tried to develop autonomous soft robots for scenarios such as the deep sea. [15]Meanwhile, multidisciplinary knowledge such as machine vision, [123,124] machine learning, [125][126][127][128] CFD, [129][130][131] and finite element method (FEM) [73] were combined to complete the modeling and design of the robotic systems.These new soft machines can operate autonomously for longer periods (Figure 4).Katzschmann et al. [80] developed the third-generation cableless underwater soft robot fish (SoFi) equipped with an acoustic control module based on the two previous generations (Figure 4a).The robot fish is hydraulically driven and can move rapidly under the commands of divers to achieve multiple locomotion modes (e.g., straight line, turning, vertical dive, etc.).It is equipped with a fisheye camera that can continuously follow and record aquatic life from 0 to 18 m deep.It achieved the first real-time interactive function for ocean exploration.
Chen et al. [132] designed a cableless soft-body swimming machine (OCTOBOT) that is preprogrammed to generate chemical reactions leading to directional propulsion functions (Figure 4b).The robot has no batteries and onboard power devices.It is propelled by shape-memory polymer (SMP) muscles.And the locomotion of the robot is achieved by using actuators that harness the large displacements of bistable elements triggered by surrounding temperature changes.Besides, it is manufactured by various state-of-the-art methods including photolithography, molding, and 3D printing.
Anike et al. [49] designed a jellyfish-like soft robot (FludoJelly) enabled by inflatable soft pneumatic composite (SPC) using inflatable structures and bionic flapping wings in nature (Figure 4c).Differing from previous fish-like robots that move in the horizontal direction, this robot was also able to move vertically and rapidly at 160 mm s À1 in the test environment while carrying a self-weight of 100 g.
Hui et al. [66] designed an eel-inspired cableless swimming robot that was hydraulically driven and fabricated by soft photolithography (Figure 4d).They used four fiber-reinforced soft FEAs to form anguilliform body waves.Soft photolithography and 3D printing technologies were employed in the fabrication of FEAs.Although both the eel-inspired robot and the snakeinspired robot invented by Onal et al. [28] adopted four FEAs, by integrating them into a serial array, these robots could still produce smooth and continuous deformation with simplified control.
higher locomotion efficiency and autonomy (Figure 4e).The system was designed based on conventional redox fluid flow batteries (RFPs) and was wrapped in soft silicone forming a bendable battery cell.Also, a hydraulic liquid actuator was embedded in the fish body and the battery electrodes were innovatively distributed over a large area of fins to increase power density.
Fan et al. [113] designed a swimming robot that looks extremely alike a frog and could be operated remotely using a radio signal (Figure 4f ).In addition, the frog-inspired robot used an articulated pneumatic soft actuator, making it small and light.This robot is a good representative of the appropriate combination of rigid and soft material.
Robert Baines et al. [8] designed an amphibious robotic turtle (ART) in the inspiration from terrestrial and aquatic turtle (Figure 4g).ART combines bioinspired aquatic, terrestrial, and transitional locomotion modes through traditional robotic components, and functional shapes through variable stiffness composites to reduce morphological and behavioral compromises.
The capabilities of robots can be improved if soft materials can be utilized in the framework of traditional rigid mechanics.Representative underwater soft robots in aspects of multilocomotion patterns, long-endurance self-powered systems, and multiple advanced technologies from 2017 to 2022: a) third-generation acoustically controlled soft robot fish (SoFi).Reproduced with permission. [80]Copyright 2018, AAAS; b) cableless soft swimming robot driven by SMP muscles (OCTOBOT).Reproduced under the terms of a PNAS license. [132]Copyright 2018, The Authors, published by National Academy of Sciences; c) inflatable SPC actuator-based soft jellyfish-like robot (FludoJelly).Reproduced under the terms of a CC-BY license. [49]Copyright 2019, The Authors, published by MDPI; d) fiber-reinforced soft FEA-based underwater soft robot of imitation eel.Reproduced with permission. [66]Copyright 2020, Mary Ann Liebert, Inc.; e) underwater soft robot based on high energy density and power density energy storage system.Reproduced with permission. [133]Copyright 2019, Springer Nature; f ) frog-inspired swimming robot based on articulated pneumatic soft actuators.Reproduced with permission. [113]Copyright 2020, Mary Ann Liebert, Inc.; and g) ART merged specialized morphogenic features for aquatic and terrestrial locomotion.Reproduced with permission. [8]Copyright 2022, Springer Nature.
The most representative underwater soft robot research is the wireless self-powered soft robot (Figure 5) developed by Li et al. [15] from Zhejiang University, which was successfully demonstrated for 10 900 m deep-sea exploration.The authors have experimentally proved the excellent pressure resistance and swimming performance of the robot.Facing the technical difficulty that conventional deep-sea swimming robots encounter such as bulky and hard containers, the innovative pressure-elastic electronic components were embedded in a soft silicone matrix with distributed precision electronic components.The locomotion of fin flapping was achieved by adding a small voltage to the DE.

Soft Materials
In contrast to rigid robots, most materials used for underwater soft robotic systems are intrinsically soft and/or extensible materials (e.g., silicone rubbers) that can deform and absorb much of the energy arising from a collision.Elastomeric materials are the major representatives adopted in the fabrication of soft robots; they own excellent high temperature and chemical resistance, excellent mechanical and electrical properties, and biocompatibility.In addition, they can maintain a certain degree of elasticity, resilience, and surface hardness in hightemperature environments above 200 ∘ C, and there is no significant variation in mechanical properties.However, the key challenge for creating soft machines which can achieve their full potential is the development of controllable soft bodies using materials that could integrate sensors, actuators, and computation all together, and thus enabling the body to deliver the desired behavior. [38]In recent years, various intelligent material fabrication techniques with low noise, high efficiency, and high flexibility have been rapidly developed and widely used in the field of robotics development, such as IPMC, [134] shape-memory alloys (SMAs), [135] SMPs, [136] PZT, [137] DE, [138] and liquid crystalline elastomers (LCEs). [139]Since actuators for existing soft robots are normally designed based on different principles and made of different soft materials, leading to varying actuating characteristics and performances accordingly.In this section, we will compare the properties of these six types of intelligent materials from both quantitative and qualitative perspectives in Table 4, and summarize other intelligent materials developed by researchers in recent years as well.

IPMC
IPMC consists of an electrolyte-swollen polymer film (typically 100-300 μm thick) sandwiched between two thin metal layers.It has a uniform distribution of anions and cations in the polymer electrolyte in the absence of applied voltage.And when bias pressure is applied to the electrodes, the cations migrate toward the cathode and the anions migrate toward the anode (Figure 6a), leading to an inhomogeneous expansion, bending, and deformation of the entire structure in a positive direction.Depending on the applied polarity, the device can be bended in both directions. [140]PMC is composed of polymer and electrodes.As one of the electroactive polymers (EAPs) that exhibits a large variation in performance under electrical stimulation, IPMC is capable of operating in both the air and underwater environments, insensitive to magnetic fields and easy to fabricate; hence, it is a good alternative of traditional actuation and sensing materials.The highest locomotion speed of soft robots can be gained by adjusting the voltage frequency applied to IPMC.Most research progress of IPMC research was achieved during the last 30 years.Researchers started to develop IPMC in late 1980s; the EAP characteristics and its electromechanical coupling properties with IPMC were discovered gradually, and then researchers continued working on its control [141] since the late 1990s.IPMCs are considered to be one of the most promising intelligent materials because of their lightweight and large displacement output under low working voltage.They can display best performance when operated in humid environment, and can be made as selfcontained encapsulated actuators operating in dry environment as well, indicating its widespread applications.

SMA
Since the advent of shape memory technology (SMT) in the 1990s, [142] SMAs were started to be applied in various industries.) is poured into the 3D-printed mold to encapsulate the electronics.Reproduced with permission. [15]Copyright 2021, Springer Nature.
SMA is the metallic material with shape-memory effect and a combination of sensing and actuating functions.So far, more than 50 alloys with shape memory effects have been found.Shape-memory effect is a phenomenon, in which a material recovers to its original size and shape when heated above a certain characteristic transformation temperature.
SMA exists in the austenite phase at high temperatures and the martensite phase at low temperatures.By heating and cooling SMA, the stretching and shrinking of SMA can be achieved (Figure 6b); thus, the actuation force for soft robots could be generated based on this principle.Therefore, SMA is widely used in manufacturing actuators for underwater soft robotic systems. [143]ere are different types of SMAs, but the most widely used one is nickel-titanium (Ni-Ti) alloy.Because of the good flexibility, it is a suitable alternative to the biological muscle.
It is noted that each individual type of abovementioned materials has its own drawbacks.Although the IPMC material can generate large bending deformation based on its ring structure through small voltage excitation, it owns disadvantages like low coupling efficiency and poor actuation speed.Despite that SMAs are light in weight and can be miniaturized, the driving temperature is difficult to control, and the drive frequency is low.Also, PZT requires a large driving voltage and is limited in small strain tolerance.

DE
DEs are one of the most extensively researched materials in the field of electronic EAPs, known for producing substantial electrical deformation, high energy density, excellent flexibility, fast response time, large driving force, and high mechanical energy density.On the downside, DE drives require prestressing and high voltage for operation.By prestretching the DE, electromechanical instability can be suppressed and a higher actuation voltage can be applied.However, prestretching will limit adaptability and flexibility.It is difficult to operate the DE actuator at the Reproduced with permission. [329]Copyright 2013, IOP Publishing; b) principle of SMA.Reproduced with permission. [330]Copyright 2016, Elsevier; c) principle of PZT.Reproduced with permission. [47]Copyright 2012, Springer Nature; d) principle of SMP's shape-memory cycle.Reproduced with permission. [136]Copyright 2021, Wiley-VCH; e) principle of DE.Reproduced with permission. [155]Copyright 2013, Wiley-VCH; f ) principle of LC transition.Reproduced with permission. [156]opyright 2010, Wiley-VCH; and g) principle of hydrogel.Reproduced under a CC-BY license. [172]Copyright 2020, The Authors, published by Elsevier.
practical voltage, and it is necessary to reduce the operating voltage by increasing the dielectric constant and decreasing the film thickness.
DEs are commonly composed of acrylic acid, silicone rubber, and polyurethane.These materials are used to enhance the mechanical, electrical, and actuation properties of composite materials. [154]As shown in Figure 6e, DEs are situated between two compliant electrodes.Upon activation, an electric field is applied between the electrodes, and the coulomb force between them extrudes the material, causing it to expand in the electrode plane. [155]The Maxwell stress caused by the change of the dielectric inside is the main reason for the deformation of the DE.

LCE
LCEs are a unique type of material that combines the properties of polymeric elastomers (entropy elasticity) with liquid crystals (self-organization). [156] By controlling the orientation of the mesogens unit of the LCEs, it is possible to realize reversible deformation.The actuating principle of LCEs is mainly based on its phase transition between ordered-disordered, as shown in Figure 6g.There are two methods to synthesize LCEs, one involving the incorporation of mesogens as side chains in siloxane elastomers, while the other involves integrating mesogens within the main chains via chain-extending reactions. [157]CEs have become a widely used active soft substance due to various remarkable properties such as soft elasticity, [158] piezoelectricity, [159] ferroelectricity, [160] and orientation. [161]Most LCEs are sensitive to temperature [162] and some LCEs are also responsive to light, [163] electrical, [164] and chemical stimuli. [165]CEs have reversible bidirectional shape memory and can be remotely triggered for large stroke, fast response, and highly repeatable actuations.Although the LC phase transition and temperature-dependent polymer chain conformational changes are relatively slower and less energetically efficient compared to natural muscle, [166] it can still lead to muscle-like, large, reversible macroscopic shape deformations.Therefore, LCE is a very promising material for research on stimulus-responsive polymer-based artificial muscles. [167]

Other Intelligent Materials
In addition to abovementioned seven intelligent materials (IPMC, SMA, SMP, PZT, DE, and LCE) that have been studied and applied to various fields, targeting specific requirements of underwater soft robots, several researchers have developed other novel materials with superior performance, such as hydrogel, [168] shape-memory gel (SMG), [169] LM-embedded elastomer (LMEE) composites, [170] SPCs, [49] shape-memory polymer composites (SMPCs), [136] azobenzene chromophore polymer materials, [60] multiple shape-memory ion-exchange polymer-metal composite (MSM-IPMC), [53] and macrofiber composites (MFCs), [171] etc. Hydrogels have a low modulus (1 Pa-1 MPa) similar to that of human tissues, and the stimulus sources for deformation are generally mild, making hydrogels play an important role in the field of soft robots. [172]A hydrogel is composed of ample water and a hydrophilic polymer network (Figure 6g).Most polymer gel artificial muscle systems absorb water and swell, while evaporation of water under dry conditions affects the stability of the material for application. [173]Thanks to the advances in soft materials and flexible electronics, great progress has been made in the development of bioinspired underwater soft robots. [174]dmittedly, existing intelligent soft materials such as IPMC, SMA, SMP, PZT, DE, LCE, and hydrogel still have defects in stress, strain, lifespan, price, etc., which cannot meet the requirements and rapid development speed of soft robots.Therefore, researchers still have a lot of improvement space in developing novel intelligent soft materials.After being stimulated by an electric current, soft robots can display more complicated locomotion/gaits, such as bending, extending, twisting, shrinking, etc., with more flexible response and better performance.Besides, the use of self-powered synthetic bionic muscles could also enhance robots' locomotion as freely as human beings.In all, the combination of synthetic materials and biological materials, and the utilization of synthetic cells made of soft electronic materials and natural muscle tissue are expected to enhancing performances of existing soft materials. [175]

Fabrication Methods
Rigid robots are generally assembled from traditional mechanical parts, while soft robots are mainly composed of intelligent soft materials.Compared with fabrication methods of traditional mechanical parts, the fabrication methods of intelligent soft materials are more demanding and require more complicated processes, deciding the properties of fabricated intelligent soft materials and the further manufacturing of soft robots.
Among these technologies, SDM, soft photolithography, and 3D printing are well-developed technologies that have been widely utilized for soft robotics manufacturing over the last 20 years.The most commonly utilized SDM was developed for rapid prototyping manufacturing (RPM) of rigid materials in the late 1990, [186] and was first used in bionic structure design of soft, viscoelastic material integration in 2000. [187]Soft photolithography is a layer-based deposition method for microstructure fabrication so that various parts can be embedded during the fabrication process, [188] and the commonly utilized materials are elastomeric polydimethylsiloxanes (PDMS). [189]However, compared to conventional photolithography methods, soft photolithography is more flexible and not restricted by light scattering, but it still suffers from deformation problem and slow speed as well.
Compared with subtractive manufacturing (SM) methods like photolithography, the additive manufacturing (AM) technique is more suitable for fabricating soft robots considering their structural characteristics, so 3D printing technology was adopted as a main fabrication method. [190]Techniques such as direct inkjet curing, [191] photopolymerization, [192] selective laser sintering (SLS), [193] and multijet fusion (MJF) [194] have been incorporated into 3D-printing manufacturing of polymers.At present, 3D printing is a typical technique for manufacturing soft robots; it is capable of building complex structures with a relatively simple manner.3D printing can also be applied to the molding of various materials. [195]In the future, 3D printing technology could be employed to efficiently fabricating the entire soft robot including sensors, actuators, and control systems by combining hard, soft, elastic, and conductive materials. [196]It is expected to serve as a complementary or even alternative approach to soft lithography.
In recent years, researchers discovered that certain 3D printed structures may change in shape, properties, and function over time when exposed to predetermined stimuli (e.g., heat, water, light, pH, etc.).Derived from such observations, they developed 4D printing technology based on 3D printing using intelligent polymers as basis and have successfully incorporated the additional dimension of time.4D printing is essentially a subset of 3D printing and owns all its advantages.The fabrication of 3D objects equipped with locomotion ability is called 4D printing, which first appeared in 2013.With the advances of this novel concept, the application scope of 4D printing has been extended to self-assembly, [197] self-adaptation, [198] and self-healing. [199]urrently, there is a wide range of applicable 4D-printing methods, including direct inkjet curing, [200] fused deposition modeling, [201] stereolithography, [202] laser-assisted bioprinting, [203] and selective laser sintering. [204]However, 4D printing is still in its infancy, and the 4D printing of soft robots relies heavily on the selection of SMMs, which might limit the performance of the robots.207] The methods mentioned above do not allow reproducible fabrication of soft fluid actuators without weakening the seams and integrating the functional structure.In contrast, the lost wax method [65] for monolithic casting allows the fabrication of soft actuators with complex internal chambers and with no joints that compromise structural integrity.It allows the formation of more complex shapes with higher dimensional accuracy than casting.Thus, soft robots can achieve different forms of locomotion by designing internal chambers of silicone rubber.In addition, to achieve continuous antagonistic movements, they can also mimic the structure and arrangement of the circular and longitudinal muscles of biological organisms. [80] Actuation, Locomotion Pattern and Power Storage 4.1.Actuation RoboTuna 1, [208] the world's first-generation tuna-inspired robots developed by David Barrett's team at MIT, and RoboTuna 2, [209] the second generation, both adopted a traditional motor-driven method.Comparing RoboTuna 2 with its contemporary's robots powered by propellers, this former is more flexible and consumes less energy.However, neither its swimming speed, instantaneous acceleration, nor flexibility is comparable with the real tuna.Therefore, the team improved the control algorithm and tail swing based on the first generation to design RoboTuna 2.
RoboPike, another robotic fish developed by the Draper lab from MIT, mimics a barracuda, and can achieve an acceleration up to 12 g with enhanced maneuverability as well; the researchers have realized the surfacing and diving functions inspired by the design of RoboTuna robot fish. [210]However, the motorbased actuation requires numerous mechanical transmission mechanisms, which results in large and complicated structure of the robotic fish, as well as high noise problem during locomotion.Therefore, bionic researchers decided to solve the problems by developing state-of-the-art actuation methods.
On the basis of the previous motor-driven robots, the emergence of pneumatic-hydraulic actuating technology and novel intelligent materials has broadened researchers' visions for the development of next-generation actuation system for soft robots.Nowadays, actuation systems for soft robots can be divided into the following three types: variable-length tendons, FEAs, and EAPs.Variable-length tendon-based methods can be further subdivided into tension cable and semiconductor memory alloy (SMA)-based approaches, in which SMAs can be embedded into the soft robots to drive soft materials like PDMS.The fluidic elastomer actuation system can be further subdivided into pneumatic and hydraulic types, and EAP can be further classified into electronic and ionic types.Each method has its unique advantages over the others and has demonstrated its significant performance in specific scenario.The main advantages and disadvantages of each actuation method are summarized in Table 5.
Fluid elastomer actuators (FEAs) are the new-generation low-power-consumption soft actuators with high scalability and adaptability and they can be driven either pneumatically or hydraulically.Feng et al. [66] adopted soft photolithography method to fabricate fiber-reinforced FEAs and designed an eel-like cableless soft swimming robot.However, in contrast to solid material actuators, these fluid actuators are not reusable; there is also a risk of leakage and the possibility of mixing in an undesirable manner when operating with solid structures and/or other fluids.Besides, fluids with small quantities are prone to evaporation problems [211] as well.
Among all FEAs, pneumatic actuators with Pneu-Net (PN) structure are the most representative ones.The PNs are composed of extensible cavities.And a certain amount of the cavities could be expanded when pressurized, while the remaining cavities couldn't be expanded.This structural feature induces the bending motion of elastic material.Compared with the method proposed by Robert F. Shepherd et al. that used chemical reaction (methane combustion) to generate explosive pressure to rapidly drive an inflatable net, the PN actuation method is limited in the gas filling speed (along the pipeline) and the widening of microchannel due to the viscosity of air.However, relevant design and fabrication can be achieved easily and inexpensively using the well-developed soft photolithography technique used for fabricating microfluidic systems.Besides, the PN actuation has another advantage of simple control as the gait or shape of soft robots can be easily modulated by adjusting the injected air pressure according to the intrinsic structural stiffness of adopted soft materials.According to abovementioned superior characteristic, researchers developed pneumatic quadrupedal robots [13] (mimicking worms and starfish), which were entirely made of elastic polymer materials.
EAPs have many types, including dielectric elastomer actuators (DEAs), ionic polymer-metal composites (IPMCs) actuators, and stimuli-responsive (SR) hydrogel actuators. [38]EAPs could also be divided into electronic EAPs and ionic EAPs.DEAs, a group of electronic EAPs, are a kind of intelligent material that can generate large strains under an electric field.Besides, IPMCs and hydrogel actuators can be classified as ionic EAP actuators.IPMCs require low actuating voltages (≤ 5 V), and currently has found a wide range of applications in bionic robotics, where conventional IPMCs depend on aqueous solutions and are mainly used in underwater robots. [67]Based on this feature, Shen et al. proposed a soft MSM-IPMC actuator.By controlling the external electrical and thermal inputs, MSM-IPMC synthetics can perform complex twisting and oscillation. [53]Hydrogels are physically and/or chemically cross-linked 3D polymeric networks that can absorb a large amount of water but do not dissolve in water, and can undergo volume expansion/contraction upon swelling/deswelling, and SR hydrogels can even undergo sharp volume changes upon the change of external stimuli like pH, temperature, salts, light, electric signal, solvent, or molecules. [212]ence, they are ideal candidates for soft actuators and are commonly used by researchers.For example, a new light-driven composite hydrogel and fabricated JMSR was developed by Yin et al. [56] JMSR is a jellyfish-like miniature swimming soft robot made of a composite temperature-sensitive hydrogelpoly(N-isopropylacrylamide)/carbon nanotubes (PNIPAM/ CNTs) and can accomplish walking and jumping gaits.In another category, electronic EAPs include all-organic composites (AOCs), dielectric EAPs (DEAP), electrostrictive graft elastomers (ESGEs), electrostrictive films (ESPs), electroviscoelastic polymers (EVEMs), and LCEs.The LCE can reversibly deform in response to external stimuli with dramatic contraction (up to 400%) and stress comparable to natural muscles.Therefore, LCEs have been considered as artificial muscles with great potential for creating a biomimetic soft robot.Palagi et al. [213] adopted LCEs to realize continuum yet selectively addressable artificial microswimmers that could generate traveling-wave self-propelling motions without additional need of external forces or torques, as well as microrobots capable of versatile locomotion behaviors on demand.
Moreover, soft microrobots are capable of performing complex tasks under the control of highly configurable artificial intelligence systems, such as delivering drugs to a person's body.The driving methods of microsoft robots include electrically responsive, magnetically responsive, thermally responsive, light responsive, and pressure-powered. [211]For example, Cunha et al. [214] designed a light-responsive LC-based soft robot active in air and aqueous environments.This comprehensive study can also be used to fabricate primitive amphibious soft robot walkers.Actuations made of intelligent soft active materials such as hydrogels and LCEs, which exhibit SR behaviors, represent a potential route toward advanced biomimetic microrobots. [213]

Locomotion Pattern
Underwater organisms have different morphological characteristics and variant gaits; different movement patterns correspond to various energy utilization efficiency.The energy efficiency of several most representative underwater organisms' locomotion patterns is listed as follows in descending order: jumping >> swimming >> snake-like gliding > rolling > getting up/turning over > inching > limb gait > earthworm wriggling, [215] among which swimming is the most common and energetically efficient pattern.Therefore, it is quite essential for an in-depth investigation on the gestures of swimming soft robots and a classification of the swimming patterns in terms of hydrodynamically driven, rowing, wave undulation, and jet propulsion.Overall, the longer the fish, the more it relies on the force generated by the trunk part, and the shorter the fish, the more it relies on the force generated by the tail fin and other forms of thrust.Hydrodynamically driven swimming is the primary gait of robotic fish.Underwater creatures that possess fins move forward by the hydrodynamic force generated when they slap their fins.For example, the electronic fish designed by Li et al. from Zhejiang University was inspired by manta rays and takes DE-film (3 M VHB) as muscle, silicone film as fin, and silicone frame as elastomer.When a periodic voltage is applied, the fins flap and the fish can generate thrust by periodically slapping its pectoral fins like a ray. [55]Besides, Breder first categorized fish locomotion into two propulsion patterns: body and caudal fin (BCF) propulsion pattern, and media and paired fin (MPF) propulsion pattern, according to the fins' location used for propulsion. [216]owing swimming is powered by the reverse driving force generated by backward stroke.Animals (e.g., octopuses, frogs, and jellyfish) can use their flexible limbs as oars to gain forward momentum by pushing water backward, such as the OCTOPUS [17] which uses eight silicone arms to carry itself forward in the water.This locomotion pattern is similar with the hydrodynamic one.But all rowing swimming animals actively push the water backward instead of simply deflecting the water current, and for this reason, their pushing parts tend to be at right angles to the current.Wang et al. [16] designed a dielectric elastomer (DEA)-driven soft swimming robot that mimics breaststroke, the structural deformation of its adaptive legs during unfolding and folding can be adjusted according to the direction of water by applying certain level of voltage, and it can also achieve flexible turnings.Another application was inspired by the dominant interaction between jellyfish and its surrounding fluid environment, where Ye et al. [217] simplified the movement of jellyfish as the rowing behavior of a thin film and developed a soft robotic jellyfish utilizing liquid metal coil, and achieved the soft rowing propulsion alike oblate jellyfish, for the first time, under controlled magnetic field.
Wave undulating swimming can be achieved just by robotic bodies, without the need of fins and other components; hence, it is mostly adopted by slender underwater soft robots, which can therefore mimic the periodic sinusoidal oscillations of the fishes' bodies (represented by eels in nature) and utilize hydrodynamics to propel itself forward.Taylor's theory confirms that the force driving a fish forward arises from resistance, which is opposite to the fish's locomotion direction but consistent with its velocity. [218]For example, a soft robot designed by Nguyen et al. [219] adopts four pairs of soft pneumatic actuators (mimicking eel muscles) and utilizes pulse signals to compress air, sending it sequentially to the actuators, thus generating a sinusoidal wave for the entire robot body enabling its moving forward under the water.The resistance generated by the torso part of long striped fish like eels is the most dominant, due to its smoothness and uniformity.
Jet-propelled swimming is widespread in numerous mollusks (e.g., scallops, octopus, cuttlefish, nautilus, jellyfish, squid, etc.).The first example of a jet-propelled soft-bodied robot is the jellyfish-like robot. [220]This robot consists of an umbrella-like silicone gel that is actuated by SMA to perform the inflation and ejection procedure.Although the control accuracy of locomotion pattern is not as good as abovementioned ones, it can generate strong explosive force and can directly utilize water in the surrounding environmental medium as the jet source.In addition, because jet-propelled robots require no additional auxiliary propulsion mechanisms such as fins, they are also suitable for microrobots.
Many scholars have proposed classifications of different types of underwater locomotion patterns for soft robots.For one, Won-Shik Chu et al. [47] proposed a new classification of swimming patterns.They expanded from fish to nonfish organisms and classified swimming mode as body/caudal actuation oscillatory (BCA-O), body/caudal actuation undulatory (BCA-U), median/ paired actuation oscillatory (MPA-O), median/paired actuation undulatory (MPAU), and jet propulsion (JET).This classification addresses the problem that the locomotion mechanisms of some robots differ from those of the target organisms, and improves the subdivision of fish locomotion mechanisms.Taking into account the locomotion mechanisms of nonfish organisms (e.g., shellfish and jellyfish), the classification of swimming modes is no longer limited to BCF and MPF.Also, jet propulsion is considered as an independent swimming pattern.Mark Hermes et al. [45] instead first distinguished two primary methods of moving through a fluid into crawling and swimming based on whether it interacts with the substrate.Then, they detailed the biological mechanisms that create crawling and swimming motions in nature as well as relevant bioinspired robots.Based on their proposed classification, the swimming patterns of soft robots can be primarily classified into vortex shedding, jetting, and drag-based propulsion according to the swimming patterns of living organisms in nature.More specifically, Hu Jin et al. [221] proposed that soft robots, while mimicking the structure and locomotion strategies of some animals so that they have similar appearance and locomotion patterns, would also enable the robots to have more than one function, for example, the arms of octopus could be used as grippers and also enable the robots to swim.In conclusion, recently more and more scholars have researched and developed underwater soft robots with novel locomotion patterns imitating biological forms, and many different classifications of locomotion patterns have emerged.However, all different classifications are in fact quite similar.They are all combined with bionics to achieve forward movement, either through mechanical structures or through chemical reactions to achieve a specific form of movement of the mechanism.
For different forms of underwater soft robot designs, the resulting motion effects are also different.When evaluating the motion effect of underwater soft robots, a crucial index is the motion speed.Therefore, we analyzed the operating speed (Table 6) of underwater soft robots according to the classification of different locomotion morphologies.
Swimming is the most common movement mode of underwater soft robots, and it is more in line with the bionic principles of the actual water environment.Among them, the structure of jet propulsion can produce the fastest movement effect, [222] but it is difficult to carry out smooth control.In contrast, the conventional flapping structure [223][224][225] can achieve a more balanced movement speed and controllability effect, and is a more widely used underwater soft robot design.
Compared with swimming, jumping is more common in froginspired bionic robots.Its motion is in the form of a fixed-step, progressive motion.From the perspective of the overall motion effect, this kind of underwater soft robot runs faster, but the continuity of movement is poor, which will greatly limit the accuracy of the locomotion. [91,226]The underwater soft robots with crawling and walking locomotion patterns [18,82,227,228] mimic the movement of the organisms living on the substrate of water more, which seriously limits the maximum speed of the robots; however, it ensures a better controllability.
Overall, the design of underwater soft robots needs to be carried out according to different environments and requirements such as speed and controllability.The way of swimming can achieve a better balance of speed and controllability, while other bionic forms of locomotion need further optimization research to realize the actual operating value in the online operating environment.
Additionally, most of the current underwater fish robots can only achieve horizontal movement but cannot move vertically in the water.The ways fish achieve the rising and sinking functions are ingenious.In bony fish, the contraction and expansion of the swim bladder can regulate the density of the body to rise or sink. [229]However, cartilaginous fish do not have swim bladders and can only balance up and down by swimming constantly. [230]nother technical difficulty of bionics is that the pressure inside the swim bladder of deep-sea fish is high, and once out of the water, the compressed gas is released and the swim bladder is swollen.Although some robots can achieve rising or sinking locomotion by altering the volume of gas or liquid in their bodies, [80,229,231,232] this approach requires some energy consumption to supply the pump, and is difficult to control. [233,234]The operation of traditional air or hydraulic pumps can also cause noise disturbances to the marine environment. [235]Some robots achieve ascent or sinking while also scaling with rigid cable, which provides continuous electricity to the robot and enables the pump to be powered.However, this approach has greatly restricted robots' walking distance by cable limitations. [236]nother commonly used method for underwater robots is to use the thrust of the propeller to control the vertical movement; [237,238] this method of levitation at a certain depth is achieved through the combined control of gravity and buoyancy, so the propeller must continuously operate for the entire duration of the vehicle's levitation.As with pumps, propellers can also be very energy intensive, noisy, and complicated to control, and they can be easily damaged by entanglement with marine organisms in operation. [239]Both of these challenges limit the broader applications of underwater fish robots beyond the laboratory.To address this problem, Lee et al. [14] developed the Flatfishbot inspired by the intermittent locomotion gliding of marine vertebrates.It controls partial buoyancy wirelessly by using a soft thermoelectric (TE) pneumatic actuator (TPA), which, in turn, mimics the intermittent locomotive gliding and turning in the water.Dror Kobo et al. [240] developed a small, centimeter-sized, backswimmer-inspired untethered robot (BackBot) which regulates self-buoyancy by controlling nucleation and release of microbubbles.This mechanism could facilitate the replacement Quadrupeds-inspired robot 0.37 [397]   Soft jumper 6 [226]   Crawling and walking Fluidic soft robotic snake 0.07 [18]   PATRICK 0.04 [227]   Worm-inspired robot 0.025 [82]   Crab-inspired soft robot 0.234 [228]   Flow-sensing walking robot 0.1 [316]   Amphibious soft robot 1.6 [9]   3D printed robot 0.06 [398]   Coconut octopusinspired robot 0.603 [326]   Swimming Hydrodynamically driven Madeleine 0.95 [224]   Spine-inspired robotic fish 0.78 [225]   Piezoelectric fish 0.03 [223]   SoFi 0.5 [80]   FEA-driven robotic fish 0.44 [12]   Tunabo 1.6 [313]   Pneumatic-actuated manta robot 0.67 [116]   PMC robotic cownose ray 0.033 [310]   Rowing OCTOPUS 0.27 [17]   ART 0.087 [8]   PoseiDRONE 7.964 [399]   Sea snailfish 0.45 [15]   Robotic octopus 0.26 [309]   Wave undulating DE-driven eel larva 0.0086 [64]   Pneumatic-actuated eel robot 0.36 [348]   FEA-driven eel robot 0.12 [66]   Magnetic soft robot 0.47 [400]   Jet-propelled Cephalopod-inspired SUUV 1.5 [243]   Squid-like aquatic-aerial robot 43.9 [222]   BUVMS with long-fin propulsion 0.23 [401]   Cephalopod-inspired CPHJE 1.8 [402]   RoboScallop 2.0 [325]  of traditional, physically larger buoyancy regulation systems, such as pistons and pressurized tanks, and enable miniaturization of autonomous underwater vehicles.Jiaoyi Hou et al. [239] designed a phase-change buoyancy control module (BCM) with flexible material as the shell, which has great potential for application in small deep-sea robots, especially in the buoyancy regulation of deep-sea flexible electronic fish.

Power Storage
Modern robots lack the versatile interconnected systems found in living organisms that are capable of converting other forms of energy into mechanical energy.Therefore, the existing robots cannot completely imitate the energy conversion efficiency and autonomy of living organisms, and the energy storage system is one of the primary limitations of the autonomy of robots.In the early years, Onal et al. [28] designed a soft robotic snake with its independent power supply and onboard drive, so that the underwater soft robot can move autonomously.Since then, Tolley et al. [241] designed a cableless pneumatic soft robot that can run independently for 2 h, and the power endurance of the robot has been improved.And, Aubin et al. [133] embedded a synthetic energy-intensive circulatory system into a cableless aquatic soft robot.Modeled on a redox battery, the synthetic system combines the functions of hydroelectric transport, drive, and energy storage into a single integrated design.The system geometrically increases the energy density of the robot to achieve up to 36 h of operation, greatly extending the continuous working time of the underwater robot.Furthermore, the underwater soft robot designed by Li et al. [15] of Zhejiang University and Zhijiang Laboratory pioneered the utilization of airborne lithium batteries for power supply last year.The test proved that it can continuously drive the fins for 45 min on the deep-sea lander at 10 900 meters above the sea.By mimicking the skull of a blunt-mouthed pseudo-lion fish, the robot's electronics made of flexible materials are distributed in such a way that they can withstand the extreme stresses in the deep sea.
In addition, based on the chemical properties of certain intelligent materials, some underwater soft robots have energy devices that utilize chemical energy, [68] biomass energy, [69] light energy, [56] water energy, [80] etc.These forms of power have corresponding advantages for specific application environments.For example, the energy source of the three generations of acoustically controlled soft robotic fish (SoFi) designed by Katzschmann et al. [80] is water.And the robot simplifies the underwater structure and alternately transports liquid from one chamber to another, which avoids the dramatic changes in the buoyancy of the robotic fish caused by the release of gas.

Modeling
The kinematics and dynamics of soft-robotic systems are different from those of conventional, rigid-bodied robots.As conventional kinematic and dynamic modeling methods are inapplicable, a large amount of new approaches for modeling and control of soft robots have been developed (e.g., constant curvature (CC)/piecewise constant curvature (PCC), variable curvature (VC), tendon-driven kinematics, FEM, machine learning techniques, Cosserat theory, spatial beam model approaches).Based on the research foundation of robotics, these methods carry out structural modeling and mechanism analysis of soft robots applied to underwater structures.
Also, this includes the modeling of soft-bodied swimmers using discrete elastic rod (DER) simulations [242] and the application of the Cosserat model to soft robots with cephalopod posterior hair. [243]Dynamic models of soft fish have also been proposed, combining bending beam theory with hydrodynamic and damping models, [244] combining beam theory with fluid models to modeling compliant tails, [245] and modeling the fish body as multiple compliant rigid segments with hydrodynamic forces. [246]hese works emphasize the necessary to modeling not only the deformation of the compliant robot, but also to considering the complicated coupling relations between the robot and the water.Continuum-type underwater soft robots have more complex dynamics due to their nonlinear materials, which is also combined with the multiphysical coupling properties derived from their interaction with the water environment, normally leading to large errors between simulations and real experiments.Therefore, there is an urgent need for researchers to use mathematical and theoretical methods to build efficient models to reduce this error and make controllers more reliable for broader applications.
Based on the research on the basic structure of the robot, the researchers conducted a modeling analysis, in considering of the water environment and the solid-liquid interaction between the robot.The basic approach is to simplify the problem as a coupling between objects, and the robot is regarded as a series of rigid connections or continuously deforming beams, and the fluid is regarded as a quasi-inviscid fluid. [247]To avoid complex wake modeling, Lighthill [248,249] confined the hydrodynamic analysis to a certain range around the robot while separating it from the wake by a plane perpendicular to the eel's spine.As the nonviscosity fluid environment is defined, the force exerted on the robot has pure inertial properties, which can be modeled and analyzed through the Lighthill reactive model.
For the operation of underwater soft robots, eddy currents are also an important factor.On the basis of Lightill's theoretical research, Candelier [250] conducted research on active and passive swimming of soft robots under the influence of Karman's vortex street.Through the morphological analysis of the swimming process, the influence of the energy charging and discharging process in the morphological changes of the fish body was quantified.
The most common simplification method used in kinematic modeling is the CC method.However, the PCC method is a version developed from CC that assumes the strain to be piecewise constant, with each segment of the soft structure having a constant strain. [251]Besides, based on the PCC model, there are several other approaches for the forward kinematics of their respective systems.For example, one uses Bernoulli-Euler beam mechanics to predict deformation; [252] another develops relationships between joint variables and curvature arc parameters which is applicable for high/ medium-stress robots; [253] also a group proposes a model to describe robot deformation for low-stress drives. [254]verall, the underwater soft robot model based on Lightill has achieved excellent research results and has been verified in the actual environment.Other research methods have also verified the feasibility of soft robot modeling through simulation and experiments.For underwater soft robot modeling, the online application of the model in the water environment is a further research content.After reasonable model simplification, the modeling speed of the existing model will be further improved, so that it can be applied to the online control of soft robots in the actual environment.

Sensing and Control
The motion control system is an important basis for the normal operation of the soft robot.For a complete soft robot motion control system, in addition to the structural design of the soft robot as the executive layer, the perception layer and the control layer are also important components.Therefore, it is necessary to complete the feedback of environmental information through the sensing system.Finally, the generated control output signal is transmitted to the soft robot to achieve the task goal.Table 7 presents some of the existing modeling, sensing, and control techniques for underwater soft robots.
However, the unique variability of soft robots makes many sensors and actuators for rigid robots cannot be adopted directly.As a result, researchers have developed methods to fabricate sensors using shape deposition techniques, soft lithography, 3D printing, and other techniques.These methods integrate relevant knowledge in microfluidics, 3D printing, integrated circuit design, and smart soft materials.
Although the use of soft sensors is not yet common in relatively mature underwater robots represented by AUVs and ROVs, there is an increasing demand for the development of soft robots with better compliance, safety, and high flexibility.Precise control is significant for underwater soft robots to achieve specific job requirement.In recent years, several high-performance underwater soft sensors have been developed.Yu et al. [255] proposed a novel contactless sensing mechanism by changing the electron transfer pathway.Functions like detection of underwater environmental variations, object recognition, information delivery, and even identification of human standing posture can be realized.Subad et al. [256] presented a facile and cost-effective synthesis technique of a flexible multidirectional force sensing system, which is also favorable to be utilized in underwater environments.Gul et al. [257] designed a fully 3D printed multimaterial soft bioinspired whisker sensor for underwater-induced vortex detection.An underwater strain sensor based on this antiswellable hydrogel is further developed by Ren et al. [258] to monitor the movements of underwater sports.Inspired by the lateral line of fish that can sensitively sense the water depth and environmental stimuli, an ultrathin, elastic, and adaptive underwater sensor based on Ecoflex matrix with embedded assembled graphene sheets was invented by Wang et al. [259] In addition, soft sensors are widely utilized in other fields, and could also be employed by wearable devices in the field of electronic medicine.
Rigid electronic devices have the advantages of high integration, high speed, high commercialization, and low cost.There are two modes of existing soft robots equipped with rigid electronic devices: fully incorporated into the body of the soft robot [12,50,[260][261][262] or placed externally, [28,[263][264][265][266][267][268] but admittedly, conventional rigid electronic devices are limited in their miniaturization, biointegration, and the ability to withstand tensile torsional compression.Therefore, the use of soft sensors with better performance is of great significance to improve the environmental adaptability, integration level, and comprehensive performance of underwater soft robots.
Compared with rigid robots that can be modeled using traditional rigid-body dynamics, forward kinematic modeling of soft robots is relatively difficult due to the hyperelasticity, viscousness, nonlinear large deformation, and other characteristics of the soft materials.And the infinite DOF makes the inverse kinematic modeling more difficult to realize, so the accurate control of soft robots has been a topical issue in soft robotics research.In addition, the control schemes of soft robots are not only influenced by different application scenarios, but also depend on different actuation approaches.
The control of soft robots can be divided into two categories: open-loop control (Figure 7a) and closed-loop control.For example, the FEM [10,80,115,269,270] is one of the most popular model-based open-loop control strategies; it represents soft robot (continuum with infinite DOF) as discrete elements; due to the long computation time, FEM is normally used for verification or prediction and has been used for the reverse design of robot structures.
Open-loop control based on model-based analysis is the most common method for motion control of soft robots.It can modulate the overall deformation of soft robot by switching the on/off of small actuators.Du et al. [74] improved the dynamic model (e.g., system identification) and open-loop control signal by leveraging the gradients from the differentiable simulator; based on a self-designed four-legged starfish-shaped soft swimming robot, it is experimentally confirmed that this method not only reduces the gap between simulation and reality but also improves the practical performance of the open-loop controller.However, in the presence of nonlinear disturbances, open-loop control might generate errors that cannot be self-adjusted due to the linear mapping output without feedback.
In contrast, the closed-loop control system based on the feedback system has a better ability of active disturbance rejection.Closed-loop control methods have also been demonstrated for the position control of soft robots. [271,272]Compared with open-loop control, which uses compensating input to determine the problems encountered in output, closed-loop control is more accurate.For example, Patterson et al. [227] presented a mobile and untethered underwater crawling soft robot, PATRICK, paired with a testbed that demonstrates closed-loop locomotion planning.PATRICK demonstrates the utility of designing soft robots with high-dimensional execution spaces, iterative planning of advanced motion primitives, and the use of vision-based state feedback.Closed-loop control can also be divided into first-level closed-loop control (Figure 7b) and second-level closed-loop control (Figure 7c).
Compared with the open-loop control system, one of the advantages of the closed-loop control system is that it is more robust and more suitable for the online control requirements in the actual environment.The representative one is the bionic frog-shaped robot designed by Huang [273] through DERs.Under the premise of ensuring the overall smooth operation through the streamlined structure design, the frog robot completes the closed loop of the overall control system through attitude perception.More importantly, on the basis of hardware experiments, the online control strategy output based on simulator processing is carried out, which is of great significance for the online control of underwater soft robots in actual environments.
In addition to the basic control framework, the specific controller design is also an important research content of soft robots.According to the goal-driven strategy, controller design can be divided into three levels.Low-level controllers drive the actuators, whereas mid-level ones are responsible for the kinematic and dynamic control, and high-level control involves Global vision -Global vision-based system -Offline output conversion advanced trajectory and path planning for tasks such as obstacle avoidance.Besides, state-driven strategy can be classified as static and dynamic.The widely used approaches are based on static controllers and utilize the corresponding kinematic models.However, they rely on steady-state assumptions, which limit their speed, efficiency, and accessibility.Another approach is to use dynamic controllers, which can plan and exploit the complex dynamics of the system.However, this is not simple because the complexity involved in producing an accurate dynamic model goes far beyond the development of a kinematic model.In addition, adaptive controllers based on linear models were introduced into soft robotics.[276] An octopus tentacle-like soft robot arm operating in an underwater environment produced by Xu et al. [277] was equipped with an adaptive vision servo controller.
The developed adaptive algorithm allows online correction of image distortion caused by the deflection of light through different media.Unlike on-the-ground visual servoing controller, underwater environment exerts multirefraction effects.Nowadays, control of soft robots has been further optimized as the field of artificial intelligence has been increasingly researched.Fitting an adaptive control process by machine learning (ML) based on sensory signals is one of the most widely used methods in soft robot control.Depending on the application scenario, inverse kinematic learning (IKL), forward dynamics learning (FDL), and reinforcement learning (RL) can be used to achieve complex behaviors such as pose perception and feedback operations of the soft robot itself. [278,279]In particular, RL is undoubtedly the most promising control method in robotics these days.As a subcomponent of ML, RL is the process of continuous learning by trials and errors based on feedback from the environment, and then adjusting and optimizing its state information to find the optimal strategy, or find the maximum reward.RL can be implemented regardless of whether a model of the system is known.Commonly used model-based reinforcement learning (MBRL) is policy iteration and value iteration, and commonly used model-free reinforcement learning (MFRL) is state-action-reward-state-action (SARSA) and Q-learning.The difference between them is whether the agent can predict before it executes its action.MBRL can predict the upcoming events in the virtual world and adopt the most beneficial strategy for the robotic system.Deep learning (DL), the ML with deep neural networks, fits data through neural networks consisting of multiple layers of neurons.DL does not require hand-designed rules and can maximize the potential features of the data.So it has great advantages in the field of ML.RL can also be implemented by deep neural networks.Over the last decade, deep reinforcement learning (DRL) control algorithms developed by combining DL and RL have been able to solve many complex problems and enable intelligent robotic systems with greater autonomy. [280]RL integrates the powerful understanding of DL for perceptual problems such as vision with the decision-making capabilities of RL to enable end-to-end learning.After building a soft robot that can swim using a DEA, Li et al. [127] introduced a data-based control framework to solve the subsea motion problem of a soft robot, using DRL and demonstrated the potential of using DRL to improve the motion of a mobile soft robot.In addition, soft robot fish (SOFI) is equipped with a deep Q-learning (DQL) training system by combining DL and Q-learning.
For the sensing and control of underwater robots, conventional manual control [129,220] is the most common and widely used control method.Limited by the limitations of information processing and manual control in the marine environment, this control method has low control precision and low automation level.In contrast, open-loop control [74,115,281,282] can complete the fixed operation of the robot's motion effect through systematic control output.But the lack of information feedback will lead to serious static error of control accuracy.The semiclosed-loop control strategy [283] based on information feedback goes a step further, realizing a certain degree of self-calibration and effectively improving the control accuracy of underwater soft robots.
There is no doubt that a strict closed-loop control [80,124,284] system is a more reasonable control system for the robot control system.However, limited by the design of the control system and the complex water environment, the closed-loop control system of underwater soft robots still needs further development.In addition, the accurate perception of information under underwater conditions is also an important content to be optimized.Through the development of sensing technology in the fields of acoustics, optics, and electromagnetics, a more complex sensing system will provide more sufficient information support for underwater soft robots, thereby realizing intelligent closed-loop control.

Microswimming Robots
The 21st century has witnessed an increase in the variety of robots, and robotics research has begun to enter a new area of development with the emergence of microsystems.Microrobot generally refers to the small size as well as the small workload of the robot.With the development of microelectromechanical systems (MEMS) technology, microrobot has a wide range of applications in biomedicine, aerospace, agriculture, and medicine.Among them, the miroswimming robot can move in liquid environments, and its superior body compliance and small volume allow it to enter various liquids and spaces that are inaccessible to human beings.
There are two common swimming patterns in the microscopic world.The first relies on wiggling the long, flexible paddle-like flagella of flagellates and sperm to generate propulsion for swimming. [285]The other is to spin the bacteria's spiral flagella to generate propulsive swimming. [286,287]s the size of the robot gets smaller, considering the limitations of processing methods, and electronic and energy storage device sizes, this leads to differences in manufacturing materials, actuation methods, locomotion patterns, power storage, sensing and control, and modeling of underwater soft robots under micro-and macrolevels.Soft smart materials and biohybrid materials have a transformative impact on the application of small swimmers in biomedical field. [288]In manufacturing materials, many researchers have utilized the properties of intelligent materials to convert electrical, magnetic, thermal, light, and other types of energy into the mechanical energy needed for underwater microsoft robots.For example, Huang et al. [60] utilized the light-driven properties of azobenzene chromophore to create a microswimming robot with complex motor functions.Zhang et al. adopted flexible magnetic homogeneous composite materials to produce a microsoft robot, and experimentally demonstrated the swimming ability and maneuverability of the robot, and quantitatively described its swimming performance under various conditions. [289]esides, due to the small size of the swimming microrobot, it is difficult to realize the built-in energy supply and drive of the microrobot, so the energy supply and drive is the primary problem of the microrobot to be solved.In addition, cable-free actuation can reduce the size of the robot to better meet the working environment of the micro-operation, so cable-free is an inevitable trend in the development of microrobotics.Currently, researchers have proposed the use of microorganisms, [290,291] chemical reactions, [292] electric fields, [293] magnetic fields, [122,294] and other solutions to drive and control them.However, piezoelectric actuation requires high voltages and bacterial actuation needs to keep the toxicity of the cells low, among others, and these control conditions limit the range of applications for microbots.Considering the safety and biocompatibility of in vivo or tissue-engineered environments, the external magnetic field actuation approach is more suitable for the actuation and control of cable-free microrobots.For example, a closed-loop control method of soft magnetic film swimming microrobots for 3D arbitrary path following at low Reynolds numbers by visual serving was proposed. [295]The path curve is divided into a series of line segments.Different complicated paths drawn by users through a 3D mouse without the input of parametric equations are followed by swimming robots during experiments.Multimodal collective swimming of ternary-nanocomposite-based magnetic robots capable of on-demand switching was reported. [296]The controllable collective actuation of these biomimetic nanocomposite robots can lead to versatile robotic functions, including microplastic removal, microfluidic vortex control, and transportation of pharmaceuticals.

Discussions
With the rapid development of underwater soft robotics, researchers have been developing new technologies in terms of manufacturing materials, actuation methods, locomotion patterns, power storage, sensing and control, and modeling.To summarize briefly, these relatively well-developed materials include silicone rubber, SMA, IPMC, SMP, PZT, DE, LCE, and hydrogels.The manufacturing methods include SDM, 3D/4D printing technology, soft photolithography, threadreinforced pneumatic chambers, and compressible needle casting.The actuation systems are divided into three major categories: variable-length tendons, fluid-driven, and EAP.For aquatic animals, swimming is a more common and efficient locomotion pattern; therefore, we focus on four swimming patterns: hydrodynamically driven, rowing, wave undulation, and jet propulsion as discussed in details above.Meanwhile, considering the flexibility of underwater robots, the power storage is evolving toward independent portability, miniaturization, long endurance, and energy diversification.The infinite DOF of soft robots makes inverse kinematic modeling more difficult to achieve; therefore, the accurate control of soft robots is still research hotspot.Nowadays, with the advancement of ML, soft robots with sensor-detected data for closed-loop control can improve the accuracy of control, and RL is undoubtedly the most promising ML.Moreover, modeling plays an essential role in fabricating the robot, which can facilitate researchers to optimize the design.The modeling is based on mathematical theory for soft robot nonlinear ontology modeling to minimize the error between simulation and real experimental data.Nowadays, some researchers have respectively modeled continuum robots based on CC/PCC, FEM, VC, Cosserat theory, and so on.
Underwater soft robots are mostly made of soft materials; compared with underwater rigid robots, the degree of freedom is highly redundant, more flexible movement, and can change their form actively or passively according to the surrounding environment, underwater soft robots have higher flexibility, safety and adaptability, in human-machine interaction, complex fragile products grasping and underwater work in narrow spaces have incomparable advantages, and greatly compensate for the shortcomings of rigid robots.It makes up for the shortcomings of rigid robots and pushes the design, modeling, sensing, control, and application of robots to a higher platform.Due to the similarity between soft robot ontology materials and soft biological tissues, they can walk freely in unknown terrain, [50] well simulate the locomotion modes of mollusks, [17,143] withstand huge impacts without damage, [143] and traverse complex and narrow spaces. [297]These properties break through the limits of rigid robots and greatly simplify the design of the robot's body structure and control system, providing new ideas for the research of bionic mobile robots.
Underwater soft robots involve multidisciplinary knowledge due to their special operation environment and the materials adopted in fabricating their bodies.Therefore, if we want to make a breakthrough in this field, all disciplines related to its development need to be advanced.It is still at the beginning stage of its development and many technologies are still inadequate, so it is very challenging to create robots that are as advanced as existing land robots.
First, Figure 8 represents the distribution of different disciplines of underwater soft robotics in recent years, reflecting the significance of the discipline in the form of proportion, with a higher proportion representing the higher significance of the discipline.By compiling the distribution of research disciplines in this field in the last decade, in addition to engineering and robotics, it is found that the direction of the material occupies the largest portion among others.In contrast, the mechanics direction occupies the least portion (Figure 8).Second, by compiling the research results of the researchers in the field of underwater soft robots, we found that it can be divided into two main directions.The direction I includes the research of actuators and sensors based on soft materials, and direction II focuses on studies of mechanical structures and locomotion control, which is further development of control strategies and the utilization of materials, actuators, and sensors in mechanical structures.Both the two directions are complementary to each other, and the number of publications in recent years related to each of them is shown in Figure 9. Since the 2000s, the number of articles on underwater soft robots in both directions has been continuously increasing following similar trend.Direction I has been rising more rapidly, the researchers in this direction are becoming numerous, and the achievements are more matured.With the joint efforts of those researchers, some high-performance actuators and sensors have been developed utilizing intelligent soft materials and other technologies.Therefore, there is an urgent need for improvements in direction II, to develop better control strategies and combinations of mechanical design to make underwater soft robots with better performance.
In recent years, underwater robotics has made great progress in structure, drive technology, modeling methods, control algorithms, etc.However, as a brand new kind of robot, the research of soft robotics has just started.Underwater soft robot research involves materials, chemistry, machinery, control, and other disciplines.There are still huge challenges in material selection, structure design, control, and sensing technology. 1) Materials: The realization of underwater soft robots with biomechanical intelligence similar to mollusks requires the development of new active soft materials, and the manufacture of materials with different mechanical properties in different directions and under different underwater pressures, which will bring a breakthrough in pressure resistance, cold resistance, and nonlinear motion for soft robots.2) Structure: Although underwater soft robots have good flexibility, they are negatively affected by low loads, poor stiffness, and low strength in some applications.Although it is possible to improve the stiffness of soft robots by adding materials such as elastic fibers into silicone materials or by using the jamming principle, [298] this method does not bring about a quantum change.Intelligent variable stiffness materials can be combined with rigid structures to develop underwater rigid-flexible coupled robots, which can not only ensure the flexibility of the soft robot but also increase the stiffness of the robot by several orders of magnitude under special needs, so that the robot's loads and other operational capabilities can be improved, thus realizing the rigid-flexible coupling and variable stiffness of soft robots.3) Control: The underwater soft robot has infinite DOF, and while the number of drivers, in reality, is limited, achieving accurate real-time control is very challenging work.Therefore, it is very meaningful to study the bionic intelligent control algorithms for soft robots.This requires research on the equivalent method of describing the infinite-dimensional distributed parameter model of the soft mechanism with a finitedimensional model and establishing an equivalent control model based on the optimization method by comprehensively considering the complexity of the model and the control performance of the system.In the future, algorithms combining distributed neural system control and classical control methods can be studied to achieve precise control of robot morphology and position and to establish performance control strategies for flexible mechanisms.4) Integration of actuation, sense, and control: Flexible robots mostly use intelligent materials and advanced structures and can complete the corresponding actions through preprogramming to achieve drive-body integration.With the development of embedded flexible sensors, flexible electronics technology, and 3D printing technology, the integration of drive sensing control of soft robots becomes possible.The integration of flexible sensors into the soft robot can make the robot sense more external information and achieve intelligent control.For example, the underwater soft robot with soft gripper can sense the shape of the object when grasping the object, acquire the information about the living body in the underwater rescue, and obtain the information about obstacles and target objects in the underwater cruise.However, there are still significant challenges to achieving the integration of sensing and driving of soft robots.For example, ways to achieve the integrated structure design of the flexible sensors and the body, and to improve the accuracy and response frequency of the sensors without affecting the mechanical properties of the robot body still need to be explored.
Overall, underwater soft robots are playing a broader role in underwater grasping, mobility, detection, military, and other fields.Underwater soft robots have higher safety, can actively or passively adapt to the shape of the object, and have absolute advantages in grasping objects of different shapes, especially soft and fragile objects, [299] and the realization of sense-driven integrated grasping of fragile targets underwater will further advance the development of human-robot integration technology.In addition, the underwater soft robots will develop toward the goal of faster swimming speed, more flexible movement, lighter volume, and a highly integrated system.Fast-moving underwater soft robots are more conducive to the realization of rapid underwater cruising tasks.A high degree of flexibility and miniaturization will be more conducive to the soft robot underwater narrow space detection.A high degree of system integration means a high degree of integration of structure, hardware, software, and control algorithms, which will be conducive to the realization of the intelligent operation of the underwater soft robots, and has a broad application prospect in the process of air, space, and ground cooperative operation and underwater cluster operation.

Outlook
The current research is intended to establish a unified foundation for the process of building complete and functional underwater soft robots (with superior performance, e.g., multimotion mode switching, environment sensing and interaction, low COT, autonomous cruising and maneuvering, etc.) on the same basis as traditional underwater rigid robots.After compiling seven existential challenges in Section 2 and reviewing the mentioned aspects above in recent years, we believe that the following nine future development directions are still of great investigative value in this field.1) Currently, research on intelligent soft materials is one of the hotpots in underwater soft robotics, mostly used to make actuators.However, the performance of existing intelligent materials still has a large improvement space.The applicability of underwater soft robots based on the novel intelligent material actuation technology is poor, some of them are still in the laboratory stage, and the practical application scenarios need to be further explored.2) Considering the practicality of underwater soft robots, the design of amphibious soft robots is sought for the future, which will lead to a significant improvement of the robot's adaptability and applicability.Amphibious soft robots are able to adapt to both terrestrial and underwater environments through deformation, and the more efficient the deformation, the more smooth the transformation between different locomotion patterns (e.g., swimming, walking or crawling, and floating).
3) To address the problem of underwater fluid resistance, the design of the robots' shapes and actuators can be continuously optimized by mimicking the unique biological forms in nature so that their shapes are more consistent with hydrodynamics and the influence of resistance can also be reduced.Thus, the speed and efficiency of movement can be improved, and from the energy consumption perspective, the range of movement can also be extended.4) With the increasing scarcity of global resources, deep-sea resources will become more important, and thus deep-sea resource exploration and exploitation technology will become an important area for the world in the future.We need to overcome the problems of high-water pressure variation, short energy supply range, and limited vision in the light-starved deep sea to develop robots that can dive into the 10 000 m-deep sea safely and complete specific tasks independently.5) More efficient control algorithms can be developed to lower the robots' energy losses and improve their locomotion efficiency for different intelligent materials.On this basis, after diving into the sea floor, the robots will be able to detect the environmental data of the sea floor and cooperate with specific actuators to assist human beings in oil extraction and pipeline laying.6) The designed underwater soft robots can also mimic the cluster behavior that is widespread and can collaborate in self-organized groups, anchor themselves to specific structures, crawl on the seabed, or wander and explore specific areas in groups.7) The ocean is much more complex than the laboratory environment, and the robots can be easily swept away by underwater currents during their autonomous operation.Hence, the robots designed in the laboratory need to optimize their locomotion capabilities for practical applications before they are formally put into practice.It is necessary to combine traditional control methods and develop new control algorithms suitable for the complex dynamic underwater environment and the special mechanical structure of soft robots.8) Underwater communication can be mainly developed in terms of neutrino communication technology.In the future, researchers will mainly simplify its technology and develop neutrino communication devices so that this excellent communication can be generally applied.9) To achieve closedloop control of a soft robot, sensors need to be integrated into the soft robot system to provide sensing feedback.Flexible sensors play a huge role for soft body sensing.The performance of flexible sensors pertaining to stability, selectivity, and sensitivity needs improvement.
The development directions mentioned above are multidisciplinary and require a great deal of interdisciplinary effort and collaboration.Therefore, it is necessary to combine the advantages of multidisciplinary intersection and work together from various emerging areas of different aspects such as bionics, materials, mechanics, and control to develop soft robots with multifunctional, highly integrated, and highly intelligent soft robots.It is certain that there will be further challenges as underwater soft robotic evolves.

Conclusions
Currently, most underwater soft robots are still in the laboratory stage, but their applications are very promising.In the future, the market for underwater soft robots could be divided into locomotion pattern and manipulation pattern.The locomotion-based underwater soft robots could be used for deep-sea resource explorations, underwater scientific research, military investigation, etc.And the manipulation-based underwater soft robots could be used in a series of underwater tasks such as marine oil extraction, port construction, naval construction, submarine line pipe maintenance, marine garbage salvage, etc.Therefore, the development of underwater soft robots will further provide more innovative solutions to promote the survey of the natural marine environment, the exploitation of marine-rich resources, the prevention and control of marine pollution, and the protection of marine biodiversity.They assist scientists to discover large unexplored areas of the deep ocean while human beings are limited in their capabilities.
In conclusion, this review presents an up-to-date summary of the existing underwater soft robotics, in categories including intelligent soft materials, fabrication, actuation, locomotion patterns, power storage, sensing and control, modeling, and microswimming robots, respectively.First of all, we have briefly reviewed the development trend and the research status in the field of underwater soft robots, and reviewed general underwater wireless communication methods.Meanwhile, we summarized reported works regarding biomimetic underwater soft robots, providing an overall picture for researchers who desire to have a comprehensive understanding of biomimetics.The body sections aim to provide detailed summary, classifications, comparisons, and statistics of each individual technique or component of underwater soft robots.Last but not the least, we discussed the existing challenges of underwater soft robotics and proposed the perspectives as well.
This review could serve as informative guidelines for experimentalists and practitioners in the field of underwater soft robots, and more importantly, the representative biomimetic achievements, intelligent soft materials, fabrication, actuation, locomotion patterns, power storage, sensing and control, modeling, and microswimming robots technologies of underwater soft robots in recent decades, as well as the existing challenges and future directions, are summarized in this review to promote a more comprehensive understanding in this field.Tiefeng Li received his Ph.D. degree from Biomedical Engineering, Zhejiang University, Hangzhou, China, in 2007.He was a visiting scholar at Royal Holloway, University of London, UK, in 2012 and 2014.He is currently a professor with Zhejiang University, Hangzhou.He is also a recipient of the National Outstanding Youth Fund.His research interests include soft matter mechanics, smart material and structures, soft robotics, underwater equipment, and healthcare device.

Figure 2 .
Figure 2. Representative biomimetic underwater soft robots.(Description of the functionality and performance metrics located under the picture of robots: v represents the speed with the unit of m s À1 .The unit of COT is J Nm À1 .COT cannot be calculated when any of the mass, speed, or energy is unknown.Additionally, it also presents the locomotion modes of each robots.)The images are reproduced with permission, by copyright as follows:

Figure 4 .
Figure 4. Representative underwater soft robots in aspects of multilocomotion patterns, long-endurance self-powered systems, and multiple advanced technologies from 2017 to 2022: a) third-generation acoustically controlled soft robot fish (SoFi).Reproduced with permission.[80]Copyright 2018, AAAS; b) cableless soft swimming robot driven by SMP muscles (OCTOBOT).Reproduced under the terms of a PNAS license.[132]Copyright 2018, The Authors, published by National Academy of Sciences; c) inflatable SPC actuator-based soft jellyfish-like robot (FludoJelly).Reproduced under the terms of a CC-BY license.[49]Copyright 2019, The Authors, published by MDPI; d) fiber-reinforced soft FEA-based underwater soft robot of imitation eel.Reproduced with permission.[66]Copyright 2020, Mary Ann Liebert, Inc.; e) underwater soft robot based on high energy density and power density energy storage system.Reproduced with permission.[133]Copyright 2019, Springer Nature; f ) frog-inspired swimming robot based on articulated pneumatic soft actuators.Reproduced with permission.[113]Copyright 2020, Mary Ann Liebert, Inc.; and g) ART merged specialized morphogenic features for aquatic and terrestrial locomotion.Reproduced with permission.[8]Copyright 2022, Springer Nature.

Figure 5 .
Figure 5. Self-powered soft robot in the Mariana Trench: a) design and fabrication of the soft robot; b) top view of decentralized electronics; and c) the precursor (Dragon Skin 20) is poured into the 3D-printed mold to encapsulate the electronics.Reproduced with permission.[15]Copyright 2021, Springer Nature.

-
Fast response time -Sensitive to external forces -Larger drive voltage

-
High bandwidth and efficiency -Need prestressing -Difficulty in driving -Fast response time -Poor physical robustness

Figure 7 .
Figure 7. Diagram of different kinds of control: a) open-loop control diagram; b) first-level closed-loop control diagram; and c) second-level closed-loop control diagram.

Figure 8 .
Figure 8. Disciplinary distribution of underwater soft robotics in recent years.Keyword "underwater soft robots" in the Web of Science and then filter the papers in the "Core Collection," refer to the data in the "Research Areas" of these papers to produce the graph.

Figure 9 .
Figure 9. Underwater soft robots query from Google Scholar, January 2023.Keyword "Underwater soft robots AND (materials OR actuators OR sensors)" for blue input and "Underwater soft robots AND (mechanical structure OR control)" for red input.

Table 1 .
Comparison of underwater rigid robots and underwater soft robots.

Table 2 .
Performance comparison of three underwater wireless communication methods.

Table 4 .
Comparison of representative intelligent material properties from quantitative and qualitative perspectives.

Table 5 .
Advantages and disadvantages of the three actuation methods.

Table 6 .
Taxonomy and speed of underwater soft robotic locomotion patterns.

Table 7 .
Sensing, modeling, and control of underwater soft robots.