Comparative study of robotic artificial actuators and biological muscle

Biological muscles exhibit a high level of integration, in which actuators, sensors and transmission elements can be included in one component. Artificial muscles or actuators refer to intelligent stimuli-responsive materials that could reversibly deform with the trigger of various external stimuli. These materials, which have attracted tremendous attention, produce natural muscle-like actuation performance and show promising applications in robotics. After an introduction of various actuator technologies that contribute to robotic applications, a comparative analysis of the main actuation parameter is provided. The comprehensive comparisons of each kind of artificial muscle are summarised, and the promising properties that are required in robotics are presented, which highlight the development of their actuation performances and the challenges that limit their further employments. Future developmental prospects and perspectives of artificial actuators are discussed.


Introduction
Actuators have an increasingly important role in robotics. Considerable development in industry and scientific research has occurred due to the remarkable progress and achievements in actuator technologies over the past decades. Actuators have been largely applied to robots such as walking machines, humanoid robots, manipulators and biomedical devices. 1 Traditional actuators, such as electric motors, are extensively applied for robotics that rely on propulsion, rotation, lifting and positioning and feature high specific power. The maximum power per unit mass of electric motors that use permanent magnets is 300 W/kg, which is twice higher when actively cooled. 2,3 Hydraulic and pneumatic cylinder actuators, which enable inherent compliance, have also been proposed. These actuators exhibit simplicity, compactness and high compressibility. [3][4][5] Regarding the power/weight ratio, hydraulic actuators have maximum typical values of 1.6-2 kW/kg, which exceed those of electric motors, with a magnitude of 100 W/kg. 6 These types of actuators enable robots to perform diversified and powerful tasks such as handling heavy weights, accurately positioning themselves, running, jumping and even performing backflips. 2,7 Although great progress has been achieved, a large gap between conventional robots and natural organisms still remains. 8,9 Many challenges, such as challenges pertaining to the materials, bio-inspired design and system integration, have appeared during the development of bio-inspired robots. One main problem is how the actuators and their topologies enable seamless interactions between human beings and robots and between robots and the physical world. 10 To understand the versatile functionalities and compliant interactions of biological mechanisms, the differences between conventional actuators and biological muscle (actuators of animals) must be distinguished.
Conventional robots are based on rigid materials for actuation and load bearing, which means that humanrobot and robot-environment interactions must be monitored and precisely controlled to avoid safety hazards to humans or the robot. 11,12 In contrast to conventional actuators, biological muscles feature compliance, which enables the load distribution on the muscle to be spread over a larger area and increases the contact time and decreases the maximum impact stresses. The compliance originates from the viscoelastic material properties of biological muscles, including the ability to change their moduli, which enables energy dissipation and the maintenance of stability during dynamic loading. Furthermore, biological muscles can freely vary their stiffness, which enables them to effectively adjust their mechanical compliance to interact safely with humans and adaptively interact with varying physical environments. 13 Typically, conventional robots consist of numerous rigid links, which are connected by separate hinge joints. 14 Each joint is actuated and controlled independently by separate hardware, including motors, sensors and controllers. When a fixed trajectory is given to an end effector, the angle of each joint required to follow this path is computed by the inverse kinematic method. In addition to the contact force between the end effector and the human/environment, the output torque of each joint is calculated by the inverse dynamic method for different situations in real time. 15 The strategy of biological muscles is different, in which actuation, transmission elements, proprioceptive sensing and control circuit and viscoelastic properties are integrated into one material architecture. 16 Rather than controlling the holistic impedance through every joint output torque as controlled in conventional robots, animals utilise the morphology, material and constraints of the entire muscle structure to adapt the stiffness and damping to an unknown environment. This control concept is also known as morphological computation. 17 Morphological computation integrates embodied intelligence into robots, which reduces the large amount of dynamic computation during real-time locomotion. Therefore, differences in the actuation principle between conventional robot actuators and biological muscles still exist. To bridge the gap between conventional robots and biological systems (human and animals), the first and primary step is to select an appropriate actuation material that can achieve biological muscle performance. The 'artificial actuators' that are extensively employed in robotics have been developed to realise a light weight, miniaturisation, flexibility and can mimic biological muscles, which offers the opportunity to improve robots. 2 This article identifies the artificial actuators that can perform as biological muscle. First, non-conventional artificial actuators that are extensively applied in robotics, including pneumatic artificial muscles (PAMs), shape memory alloys (SMAs), electroactive polymers (EAPs) and newly developed actuators. Conventional actuators, such as electromagnetic motors and hydraulic actuators, and infrequently employed artificial actuators, such as piezoelectric actuators, shape memory polymers and hydrogel actuators (HGA), have been reviewed or compared with muscles elsewhere. [18][19][20] Then, the basic working principles, classifications and applications of the extensively applied artificial actuator technologies are introduced in section 'Artificial actuator technologies'. In section 'Results', representative actuators of each technology are discussed and compared based on their actuation parameters. On one hand, the same parameters of these actuators are intercompared; on the other hand, the comprehensive actuation performances of these actuators are individually compared to that of biological muscle. The comparison results reveal the advantages and disadvantages of the introduced actuator technologies. Options for actuation are provided to satisfy different robotic application requirements, and suggestions for selecting suitable robotic actuators are offered. In section 'Future directions', the prospects of artificial muscles are proposed based on generally employed artificial actuators, which provide insights into future biomimetic design.

Artificial actuator technologies
In this part, four main artificial actuator technologies -PAMs, SMAs and EAPs -and newly developed actuators will be introduced. This study mainly discusses actuators that can be employed as actuators in robots, while other actuators fall into the categories of PAMs, SMAs and EAPs or newly developed actuators are omitted. flexible membrane. PAMs are actuated by the pressure difference between the inside gas and ambient gas, and two actuation methods can be applied: overpressure and underpressure. Usually, PAMs operate at an overpressure: as the membrane is inflated, the PAMs axially contract and exert a force that pulls the load at the end fitting (Figure 1(a)). 21 Classifications and applications: Based on the tension-carrying structure of the muscle, either embracing the membrane or being embedded in the membrane, PAMs can be classified as embracing types and embedded types. Embracing PAMs can be further subdivided into braided PAMs and plated/netted PAMs based on how the membrane radially expands, by stretching the membrane materials or rearranging the membrane's surface when inflated.
Braided PAMs. Braided PAMs are usually cylindrical in shape and composed of a cylindrical tube or bladder surrounded by a braided sleeving with a constant pitch angle (Figure 1(b)). 21 The McKibben PAM and sleeved bladder PAM are two basic types of braided muscle, in which the inner bladder is either attached to the end fitting or not attached to the end fitting, respectively. 28,29 The McKibben PAM is the most extensively applied artificial muscle in anthropomorphic robots with low cost (Figure 1(c)) and can be easily assembled in prostheses and orthotics due to the simple design. Its lengthload curves between this type of PAM and biological muscle are similar. 22,30,31 Plated PAMs. Plated PAMs have a rearranging-type membrane, which means that the membrane does not stretch when inflated. 32 The rearranging-type membrane has numerous pleats in the axial direction, which can expand during inflation and causes the membrane to unfold without material strain (Figure 1(d)). 23 No friction exists, and no energy is required to expand the membrane in the radial direction during the inflation process. Therefore, minimal hysteresis occurs, and a low threshold pressure is required to actuate plated PAMs. Plated PAMs can be applied in accurate and fast positioning devices and compliance actuation robots (Figure 1(e)). 24,32 Netted PAMs. Netted PAMs have a surrounding membrane with a low-density network. The low-density network consists of a mesh with relatively large holes and a tightly woven braid. This membrane is also the rearranging type due to the membrane structure. Netted PAMs include the Yarlott PAM, robotic muscle actuator (ROMAC) PAM and Kukolj Muscle. 25,33,34 A higher force and negligible hysteresis occur due to the absence of friction and membrane stretching. Thus, netted PAMs are usually employed in lifting devices (Figure 1(f)). 25 Embedded PAMs. Embedded PAMs have a rubber tube (load-carrying structure) that is embedded in their thin membranes, which can reduce the friction and threshold pressure due to the absence of friction between the rubber tube and the membrane. Embedded PAMs include the Morin PAM, Baldwin PAM, Paynter knitted PAM and Kleinwa¨chter PAM. 28,[35][36][37] In addition to linear motion, embedded PAMs, such as the PAM shown in Figure 1(g), can also achieve torsional motion, which can be applied in torsional inflated robots (Figure 1(h)). 26,27,38 SMAs Actuation principles: SMAs are metallic alloys that utilise the shape memory effect and can return to their origin shape or size when subject to a memorisation process between two transformation phases. The memorisation process that is applied to actuators is usually thermally or magnetically driven. 19 Typically, SMAs are deformed by an external force and can contract or recover to their original form when subject to stimuli such as heat or a magnetic field.
Classifications and applications: Based on the external stimuli that are applied, SMAs mainly fall into two types of actuators: thermal SMAs (TSMAs) and ferromagnetic SMAs (FSMAs).
TSMAs. TSMAs return to their original shape if heated after deformation, in which the transformation from martensite to austenite has a central role. At low temperatures, a TSMA has a twinned martensite structure, which can be easily deformed by an external force and converted into detwinned martensite. When the TSMA is heated by external or internal heating (Joule heating), it begins to transform from martensite to austenite at the austenite-start (As) temperature and finishes this transformation at the austenite-finish (Af) temperature. The austenitic phase is a highly ordered state, and the TSMA contracts and recovers to the predefined hightemperature shape. Conversely, the TSMA reverts to martensite after cooling, which starts at the martensitestart (Ms) temperature, and is completed at the martensite-finish (Mf) temperature. The TSMA is elongated by the defined external force (Figure 2(a)). [39][40][41] Therefore, the shape memory effect is only observed in most TSMAs, when an external stress is applied. This type of SMA is also known as a one-way SMA. In addition, some TSMAs display a two-way shape memory effect, which can act without an external stress. 42 Nickel-titanium (NiTi) alloys are one of the most extensively applied TSMAs and have been applied to small robots such as fish robots (Figure 2 (Figure 2(e)). [43][44][45][46] FSMAs. FSMAs demonstrate the shape memory effect in response to a magnetic field, which drives the shape change faster than relatively slow heating and cooling processes. This phenomenon is generated by the transition among variants of the tetragonal martensite phase that formed during cooling from the cubic austenite phase. In three-dimensional (3D) space, the martensite unit cell has a shorter size along one axis and a longer size along the other two axes than the cubic cell ( Figure 2(f)). 49 The martensite unit cell can be easily, magnetically and mechanically deformed along its shortest axis (Figure 2(g)). 47 Typically, a predefined force is applied in the short axis direction, which compresses the FSMA; it can return to its original shape when a magnetic field is applied orthogonally to the short axis direction. Due to the fast response and high specific power, FSMAs have also been employed in small-sized robots such as caudal peduncles for fish robots (Figure 2(h)) and flappers for flying robots. 48,50 EAPs Actuation principles: EAPs are smart materials that are capable of producing mechanical deformation in response to electrical stimulation. EAP materials exhibit some features that are beneficial for actuation such as great strain, high power density, versatility, scalability and low cost. 51 These characteristics allow EAPs to emulate biological muscle more closely as 'artificial muscles'. 52 Classifications and applications: According to their activation mechanisms, EAPs can be classified into two major groups: electronic EAPs (e-EAPs) and ionic EAPs (i-EAPs). e-EAPs are activated by electrostatic forces, which exert dimensional changes in response to an electric field. This category includes dielectric elastomer actuators (DEAs) and relaxor ferroelectric polymers (RFPs). i-EAPs are actuated by an electrically induced shift of ions. An electrolyte phase, in which ions are capable of moving, is required for i-EAPs, which is often a liquid. This technology mainly includes conducting polymers (CPs), ionic polymer metal composites (IPMCs) and carbon nanotubes (CNTs).
DEAs. DEAs can generate deformations as a result of electrostatic interactions between two electrodes. DEAs act as capacitors when an external voltage is applied; opposite charges are attracted in the electric field direction, while like charges repulse each other in the direction perpendicular to the electric field. Therefore, the Maxwell stresses exerted in the dielectric materials squeeze the DEAs along the direction of the voltage and expand them in the other two directions ( Figure  3(a)). A silicone or acrylic material often acts as the elastomer of a DEA, with a compliant electrode that is applied to achieve large deformation. For this reason, DEAs are extensively applied as actuators for robots: fish robots (Figure 3 [53][54][55][56][57][58][59][60] RFPs. A ferroelectric polymer material that is often used as an actuator is poly(vinylidene fluoride-trifluoroethylene), abbreviated P(VDF-TrFE). The backbones of P(VDF-TrFE) are highly polar due to the electronegativity of fluorine. The reversible conformational changes caused by the alignment of polar groups are utilised in actuation. 52 As shown in Figure 3(h), when an electric field is applied perpendicular to the chain, the ferroelectric that is originally in the nonpolar alpha phase  switches to the polar beta phase. 64 This switching causes elongation along the chain length and contraction perpendicular to the chain direction. A problem with ferroelectric polymers is the existence of a massive hysteresis and substantial energy dissipation during the actuation process, when a large electric field is applied to reverse the polarisation. 65 To eliminate the hysteresis behaviour, P(VDF-TrFE) can be irradiated or a small mass fraction of a bulky monomer can be added to disrupt the long-range order to form RFP. 66,67 Consequently, great reversible electrostriction is obtained in RFPs, which can be applied in millimetre-sized legged robots (Figure 3(i)) and swimming robots (Figure 3(j)). 62,63 CPs. CPs are also known as conjugated polymers. The electrochemically changing oxidation state causes charge addition to or removal from the electrolyte when a voltage is applied to the two electrodes, which is also known as doping. Dimensional changes are caused by electrochemical doping. 68 As shown in Figure 4(a), the CP expands when it is oxidised as the anions (A-) enter the electrolyte to balance the charge, and the CP contracts when it returns to the reduced state, in which anions have been removed. 69 The main advantage of CPs over e-EAPs is their low driving voltage, which typically ranges from 1-2 V to 10 V for higher speed and response. 70 This characteristic combined with high stress renders CPs attractive as actuators. For example, CPs have been applied in rotational motors ( Figure  4(b)). 71 In addition, CPs have also been employed to actuate micro-grippers (Figure 4(c)) and fish robots (Figure 4(d)). [72][73][74][75]   CNTs. A CNT is a single rolled sheet of graphite with a large surface area and high conductivity, whose size is on the micrometre scale. The main cause of actuation is electrostatic forces. As shown in Figure 4(e), CNTs act as electrodes and/or counter electrodes when they are immersed in electrolytes. 61 When a potential is applied, ions are attracted to the nanotubes, which cause charging and rearrangement of the electronic structure. The electrostatic repulsive forces between two charges on the CNT surface can elongate and expand the nanotubes. 83,84 Typically, actuation is generated in films or yarns that are composed of many nanotubes. For instance, CNT films that are often applied as bending actuators to drive a soft gripper can also be utilised as buckypaper actuators to actuate a centimetre-sized walking robot (Figure 4(f)). 76,85 CNT yarns perform as artificial muscles that can generate torsional and tensile motion (Figure 4(g)). 86 In addition to the electrical method, this type of torsional artificial muscle can be chemically and thermally driven. 77 IPMCs. An IPMC is composed of a polymer electrolyte sandwiched by two high-surface area metallic electrode layers, and the two electrodes interpenetrate the polymer electrolyte. 87,88 The polymer electrolytes that are often applied, such as Nafion, are neutralised with certain counterions, which are balanced by the electrical charge of the anions fixed to their backbone. 87 When an electric field is applied perpendicular to the IPMC layers, the mobile cations in the polyelectrolyte move towards the oppositely charged electrode (Figure 4(h)), which cause swelling at the negative side and shrinkage at the positive side to achieve bending of the entire IPMC actuator. 78,89 Numerous robotic applications that are based on IPMC actuators have been reported. First, IPMCs can be readily employed as actuators of  soft grippers due to their inherent bending motion. 90,91 Second, IPMCs are the preliminary technology for hydrodynamic propulsion such as IPMCs in manta ray robots ( Figure 4(i)) and fish robots (Figure 4(j)). 79,80 Finally, IPMCs can perform as robotic joints that move in parallel such as IPMCs in a millimetre-sized manipulator ( Figure 4(k)) and a snake robot (Figure 4(l)). 81,82 Newly developed actuators In recent years, various kinds of promising actuators have developed and been employed in robotic applications with different smart materials. Representative actuator technologies, including twisted fibre actuators (TFA), HGA and hydraulically amplified self-healing electrostatic (HASEL) actuators, will be discussed.
TFAs. Actuation principles: high-strength polymer fibres are transformed to artificial muscles by twist insertion. Twisted yarns of polymer fibres can generate linear and torsional actuation after volumetric expansion, which can be achieved via different mechanisms, including thermal expansion and volumetric expansion of absorption. The heat of thermal actuation is generated via two mechanisms: Joule heating and photothermal heating. 92 The absorption actuation employs moisture-sensitive silk fibre to exert deformation while changing the relative humidity. 93 The one-end-tethered twisted yarn responds to the structural volumetric change by untwisting at its free end to form the torsional actuator. The two-end-tethered twisted yarn can only change the diameter by translating and contracting in length instead of untwisting in the end, which forms the linear actuator.
A typical feature of this kind of actuator is the twisted fishing line (e.g. nylon 6) or sewing thread (e.g. nylon 6.6). First, this thermally driven artificial muscle can be constructed by inserting an extreme twist into the fibre. Second, the twist is further inserted to form a self-coiled yarn ( Figure 5(a)). 94 Finally, an annealing process is applied to yarn to avoid untwisting after disengaging the fibre end-fixing. The nylon artificial muscle can achieve large output contraction and generate super-high mechanical power. The excellent muscle stroke and work capacity are retentive as the fibre diameter changes by orders of magnitude. This nearinvariance of actuation performance indicates great scaling of structure and properties. Therefore, the nylon muscles can be converted to a two-ply ( Figure  5(b)) configuration and braided structure ( Figure  5(c)). 94,95 The scalable nylon artificial muscle enables a larger strain and stress output, which is applied in macroscopic applications, such as the light-weight prosthetic hand ( Figure 5(d)), 95 where the fingers are actuated by a twisted nylon artificial muscle and soft robotics, where two twisted nylon fibres are inserted into the substrate, which is represented by blue lines (Figure 5(e)). 96 HGAs. Actuation principles: hydrogel polymers with high water contents ( ø 90 wt%) can switch their sizes and shapes by the uptake and release of water to change the water content in the polymer work, and the polymer actuators perform deformation and movement in response to various environmental stimuli ( Figure  5(f)). 97 The main branch of the HGA family consists of pH-responsive hydrogels, which sensitively respond to pH changes in the ambient environment with swelling or shrinking deformation. When experiencing pH variation, the hydrogel polymer backbone will be ionised and the electrostatic repulsion is generated between adjacent polymer chains, which enables the absorption of water and causes hydrogel swelling. 102,103 Some hydrogel polymer can undergo reversible shifts in morphological and physical phases in response to temperature variations. The underlying mechanism of the thermally driven hydrogel is attributed to a solubility shift in aqueous solution. 104 These two methods are limited by the uncontrollability on/off state of actuation. Thus, a precisely controllable method is introduced. Electric-responsive hydrogel can swell/deswell in an electric field, which is combined of coulombic, electrophoretic and electro-osmotic interactions. 105,106 Electrical-responsive hydrogels are often employed in an aqueous electrolyte, and the direction migration of mobile ions in the solution occurs in an applied electric field, which causes contraction of the hydrogel. An osmotic pressure difference is then exerted due to the concentration gradient of ions. The hydrogel polymers generate swelling and deswelling according to the osmotic pressure increase and decrease, respectively.
To respond to various stimuli, in this way, the hydrogel-based actuators could perform deformation in various environments. HGAs are soft and wet materials that can interact with the internal constituents of the human body in a safe and human-friendly manner. Therefore, HGAs are raising tremendous attention and showing their potential in applications such as grippers ( Figure 5(g)), 98 artificial muscles ( Figure 5(h)) 99 and drug delivery. 107 HASEL actuators. Actuation principles: two types of HASEL actuators achieve linear actuation motion. First, the planar HASEL actuator, which has an elastomeric shell, is partially covered by a pair of opposing electrodes outside and filled with a liquid dielectric inside. 100 An electric field exists through the liquid dielectric when voltage is applied to the electrode. The resulting electrostatic Maxwell stress presses and displaces the liquid dielectric from between the electrodes  to the surrounding volume. Actuation strain increases with an increase in the input voltage, where the resulting hydraulic pressure causes the soft structure to deform into a toroidal shape and gain an elongation actuation movement, and this type of actuator could be stacked to achieve higher strain and force ( Figure  5(i)). 100 Second, a peano-HASEL actuator is constructed with a series of pouches made from a flexible and inextensible shell that is filled with a liquid dielectric. Electrodes cover a portion of each pouch on both side of the actuator. The electrostatic force causes electrodes to progressively close when a voltage is applied. The pressure forces dielectric fluid into an uncovered portion of the pouch, which causes the pouch to transition from a flat cross section towards a circular cross section. This transition causes linear contraction of the actuator as the shell is inextensible ( Figure 5(k)). 101 Furthermore, the parallel configuration obtained from the peano-HASEL actuator has greatly amplified the output force ( Figure 5(l)). 101 The HASEL actuator blends soft fluidic actuators with muscle-like performance of DEAs to achieve a variety of actuation modes without rigid frames or prestretched mechanisms. The hydraulic principles enable HASEL actuators to scale actuation output strain and force. The scalability gives HASEL actuators the capability to construct a soft gripper that can handle delicate objects ( Figure 5(j)) 100 and an artificial muscle that can lift more than 200 times their weight ( Figure 5(m)). 101 Unlike traditional solid DEAs, which would be permanently damaged due to a high electric field, the use of a liquid dielectric enables HASEL actuators to self-heal from a dielectric breakdown. The self-healing property improves the durability and stability of HASEL actuators, which demonstrates promise for applications in robotics.

Results
This study selected typical actuators that represent each type of artificial actuator technology. We collected parameters of these typical actuators, including their strain, stress, strain rate, work density, specific power and efficiency. First, this section successively compares typical actuators that are based on single actuation parameters, which demonstrates the superior actuators for each actuation property. These comparisons can serve as references in actuator selection to satisfy certain actuation parameter requirements. Second, this section simultaneously compares the six actuation parameters of the typical actuators with those of biological muscle to obtain insights into each type of actuator. The merits and insufficiencies of each typical actuator are given, and improvement methods and new ideas are proposed based on the comparison results.

Actuation parameters of representative actuators
This study reviewed the actuation parameters of skeletal muscle, the McKibben PAM, NiTi-based SMA, acrylic-based DEA, P(VDF-TrFE)-based RFP, polypyrrole (PPy)-based CP, CNT-based actuator, Nafionbased IPMC, nylon TFA, poly(vinyl alcohol)-based HGA and liquid dielectrics-based HASEL. These actuators are the most prevalent among each type of technology and act as samples to assess the actuation capacities of these classifications. The actuation performances of the biological and artificial muscles based on the six key parameters are listed in Table 1, where the maximum values of the actuation parameters are given. Although some of the maximum values were obtained from theoretical models or ideal materials, they demonstrate the potential upper limit for each actuator

Single parameter comparison
To evaluate these actuators for various purposes or different applications, single actuation parameters were compared according to the collected data.
1. Strain is expressed as the ratio of total deformation to the initial dimension of the material body along the direction in which actuation is being applied; it reflects the deformability of actuator materials. As shown in Figure 6(a), due to the low modulus and high dielectric strength of the acrylic elastomer of the DEA and special structure of HASEL, giant strains of 380% and 170%, respectively, are achieved at high applied fields. 100,139 The PAM and IPMC also exhibit relatively large strains that approximate the strain of biological muscle (20%-40%) Figure 6. Single actuation parameter comparisons between biological and artificial muscles, wherein the breaks on the vertical axes and in the bars indicate that the values above the break are higher by at least an order of magnitude than the values below (data are all from Table 1 due to their soft bladder or electrodes. The strains of the other actuators are less than 10%. 2. Stress is expressed as the ratio of the output force to the cross-sectional area normal to the actuation direction. The output force scales linearly with the cross-sectional area for all actuator materials, which implies that higher stress causes higher output force when actuators have the same size. In Figure 6(b), the SMA shows ultrahigh stress to 200 MPa, and the RFP, CP and CNT also exhibit relatively high stresses. High output stresses are rooted in their high elastic moduli (.1 GPa), while materials with relatively low elastic moduli have low output stresses such as the PAM, DEA and IPMC. 3. The strain rate is the change in strain per unit time and indicates how fast the material can expand or shrink in response to the motion commands delivered from the central controller.
Thus, a high strain rate is essential for materials or actuators when applied to robots with a high actuation bandwidth. In Figure 6(c), the e-EAPs, that is, the DEA and RFP, exhibit overwhelming superiority due to their high dielectric constants. The HASEL actuator is actuated by the mechanism of DEA and achieves a high strain rate accordingly. The PAM and SMA also display high strain rates due to pneumatically and electrical-heating-driven methods. However, ion diffusion and the transport velocity limit the strain rate of the i-EAPs, including the CP, CNT and IPMC. 4. Work density is the work that is generated in the actuator per cycle and normalised by the volume. This parameter indicates the amount of work that can be exerted by the actuator in a limited space. As shown in Figure 6(d), the SMA, DEA and CP present higher work densities than the other types of actuators. Actuators with a high work density are marked by either a large actuation strain or a high output stress. 5. Specific power is the ratio of the power that is generated by the actuator to its mass. With an increasing emphasis on the light weight and miniaturisation of working devices, a lower mass and a higher power are required. With a high work density and high strain rate, the PAM, SMA, DEA and RFP demonstrate high specific power, as shown in Figure 6(e). In addition, i-EAPs are still constrained by the low strain rates, which generate low specific power. 6. Efficiency corresponds to the ability of actuators to convert the input energy to output mechanical energy. This parameter determines the power supply volume and continuous working period when the actuator is applied in robotics. In Figure 6(f), electrically driven actuators, such as the DEA and RFP, show very high efficiencies to 80%. The PAM demonstrates an actuation efficiency that is similar to that of biological muscle, while the SMA has only a 10% efficiency due to heat dissipation. The efficiencies of the i-EAPs do not exceed 1% due to their low electromechanical coupling.
Artificial actuators have better performances than biological muscle from the aspect of single actuation parameters. However, the comprehensive actuation performances of these actuators compared with biological muscle remain unclear, and further discussion is needed.

Comprehensive actuation performance comparison
To understand the benefits and drawbacks of each type of actuator, the comprehensive actuation performances are analysed in this section. Artificial actuator technologies are individually compared with biological muscle to evaluate their actuation advantages and challenges.
PAMs. Of all the PAM technologies, the McKibben PAM is the most prevalent due to its low cost and simple and compact structure. This muscle is constructed with an elastic inner tube surrounded by an expandable braided sleeve. The braided sleeve translates the volumetric increase in the inner bladder to axial contraction, which is capable of generating a pulling force. The length and diameter of the braid change as the pitch angle changes. The muscle would contract until a pitch angle of 54.7°is attained, 38 which is the critical angle that corresponds to the maximum volume of the inner bladder and a zero contract force, and the muscle would stretch until the pitch angle decreases to its lower limit, which significantly depends on the braided fibre thickness and strand density. However, the tension force that is generated during contraction depends on the muscle diameter and pitch angle at rest, as well as the driving air pressure. The generated force decreases greatly with an increase in the initial angle to the critical angle and then the muscle produces negative forces. 140 Consequently, the range of contraction-extension (i.e. strain) and tension force (i.e. stress) is determined by the muscle size, woven pattern and driving pressure. As shown in Figure 7(a), the strain and stress of this PAM are slightly higher than those of biological muscle. The strain of the McKibben PAM is typically 25% in practice, and a 1.16 MPa output stress is achieved at an input air pressure of 5 bar. 18,38 Based on the output strain, output stress and ease of assembly, McKibben PAMs can be extensively applied in task-based robot hands and arms such as the Festo Bionic Soft Hand, Shadow Biomorphic Arm, Festo Airic's Arm and 7 degrees of freedom (7 DOF) robot arm developed by B. Tondu. [141][142][143][144] The most remarkable advantage of the McKibben PAM is its high specific power, which is derived from the light weight of the actuator materials and compressed air, which is typically in the range of 0.5-2 kW/ kg to a maximum of 10 kW/kg. [110][111][112] The work density is approximately 0.2 MJ/m 3 . 38 With the relatively fast strain rate of approximately 800%/s, the McKibben PAM performs better in these aspects than biological skeletal muscle, and thus, is suitable for biped walking robots, humanoid robots and Kangaroo jumping robots, 22,145,146 where high power actuators are required. The McKibben PAM's energy efficiency is approximately 0.32-0.49; the energy efficiency of human muscle is 0.4. 38,112 Despite the distinct positives, the McKibben PAM shows nonlinear strain-stress features, which are caused by viscoelastic properties of inner tube materials, the compressibility of air and complex behaviour of the braided configuration. The nonlinearity in combination with inherent dry friction between the sleeving strands and the inner tube causes actuation hysteresis, which increases the complexity of their model and control. Thereby, the McKibben PAM is not suitable for application to the scenario in which accurate positioning is required. The required compressor in the conventional McKibben PAM makes the pneumatic system heavy, bulky and unsuitable for use as a micro-actuator, where compact and weight minimisation configurations are desired.
SMAs. The NiTi SMA provides significant displacement as a hard material; the most prevalent method for manufacturing SMA is vacuum melting, where NiTi alloys are melted by vacuum induction melting in a graphite crucible, where the carbon content increases and TiC is mixed in the alloy. 147 In addition, SMA can be produced by many other ways such as the heat treating method, powder metallurgy method and thermal spray method. 148 The strain is typically 5%, with a maximum value of 10%. 108,112,113 This SMA exerts a large force per unit area of more than 200 MPa. Therefore, NiTi SMA is extensively employed as a micro-actuator or artificial muscle in robotics. The SMA reacts directly to an external stimulus, which significantly reduces the actuator's complexity and size. Thus, the NiTi SMA has a very high work density of 10 MJ/m 3 , which is 25 times that of DC motors, and an unmatched specific power of 50 kW/kg, which means that it can lift 100 times its weight. 112,113 As shown in Figure 7(a), the stress, work density and specific power of this SMA are the highest among all the actuators. Furthermore, the NiTi SMA can act in a fully 3D manner, which allows the actuator to extend, bend and twist, separately or simultaneously. These excellent characteristics have enabled application of the NiTi SMA in micro-walking robots, climbing robots, flying robots, jumping robots, swimming robots and hand prostheses. [43][44][45][46]149,150 The main challenge is the low response time and low operational frequency of the NiTi SMA, which stem from a relative high heat capacity and density of the material. These disadvantages caused difficulties in the rapid transfer of heat from active elements. The response time of the NiTi SMA can be reduced to several milliseconds, with a maximum strain rate of 300%/s, when a very large electrical current is used for heating. 114 However, the relatively slow cooling process still limits the actuation cycle time, even with an active or passive cooling method. The response time is also affected by the size and shape: the smaller diameter actuator heats faster due to higher resistivity and cools faster due to a higher surface-to-volume ratio. Another challenge is the low efficiency, which is approximately 10% theoretically and usually less than 5%, in practice; these values are much lower than the efficiency of biological muscle (Figure 7(a)). 108,115 Most SMA actuators are based on a spring configuration but the stress distribution is not constant over the spring cross section, which require a larger volume of material to be heated or cooled for the same output force and further reduces the actuator's efficiency. For optimal use of the NiTi SMA, a straight wire configuration is recommended due to a greater amount of work generated from less material volume.

e-EAPs
DEAs. Among the e-EAPs, the acrylic very high bond (VHB) elastomer-based DEA is discussed due to its soft material properties, which allow high strain and stress. The DEA manufacturing process can be subdivided into two stages. First, the coating stage: the coating will be applied to a layer of mixture with a controlled thickness on moulds with the desirable configurations, which are fabricated using lithography. A metal electrode will be applied to the corrugated side of the elastomer film. Second, the assembly stage: the laminated DE films are stacked or rolled to form the actuator. 151 High strains to 380% are achieved by the VHB elastomer materials when high electric fields are applied. However, such large deformation cannot be attained with simple linear configuration actuators. Topologies of actuators, such as stack or conical topologies, are applied to achieve large displacements, which typically range from 10% to 100%. 53,139,152 The stress of VHB elastomer is 7.7 MPa at a high voltage; 53,114 however, high-stress output is challenging with a simple film structure. Most DEAs utilise lateral expansion for force output; thus, the output stress can be amplified by stacking or rolling the dielectric elastomer and can be enlarged using them in bundles. As shown in Figure  7(b), the strain surpasses that of biological muscle, and the output stress is also higher. This type of DEA has been applied as an actuator for a hexapod robot and even for an arm-wrestling robot that can compete with a human. 118,153 The high strain and stress produces a high work density of 3.4 MJ/m 3 and a high specific power of 3.6 kW/kg. 53,65,114,115 The strain rate of DEAs is 450%/s due to the charges' quick response to an electric field. Furthermore, due to its simplicity and low modulus, this DEA is suitable for applications where mechanical compliancy and compact sizes are essential. Thus, DEAs have also been employed as flapping wings in flying robots, where a high actuation frequency is essential. 118 The actuation efficiency ranges from 60% to 80%; this high efficiency is primarily attributed to the direct transformation from electric energy to mechanical energy. 114,115 However, the high voltage used to drive the DEA is unacceptable in various applications and enforces an external electric source. 154,155 A large dielectric constant that corresponds to a greater capacitance is desired, which reduces the actuation electric field. Therefore, components, such as organic dipole molecules, high permittivity ceramic fillers or conductive nanofillers, are incorporated in the DEA to improve the dielectric constant. However, these methods often cause side effects such as decreased dielectric strength, reduced energy density or stiffening and agglomeration. [156][157][158] The electromechanical instability is the second challenge for the DEA. The Maxwell stress exerted by the DEA is above a certain threshold and triggers a feedback loop of film compression until it breaks down, which causes failure of the actuator. Therefore, a rigid prestretch structure is essential for the DEA to maintain tension and surpass the threshold. 139 However, the prestretch mechanism produces an additional load, which reduces the DEA work density and specific power. A stress concentration occurs at the interface between the rigid frame and the soft elastomer material.
The third obstacle is that the compliant electrodes are required for the DEA. The DEA's electrodes must be more compliant and stretchable than elastomer film to prevent a stiffness constrain. In addition, the durability and reliability of electrodes should be considered; electrodes have to maintain conductivity across several millions of actuation cycles with large deformation.
RFPs. The RFP based on P(VDF-TrFE) has a phase transition between ferroelectric transition and paraelectric transition with an increase in temperature. In the ferroelectric phase, a reversible molecular change between the nonpolar form and the polar form can exert polarisation change; this characteristic leads to output strain and high stress. Poly(vinylidene fluoride) (PVDF)-based actuators are usually fabricated via the layers that formed by gluing two PVDF sheets together; the layers were clamped to a substrate with their polarisations in opposite directions. 159 Recently, the electric poling-assisted additive manufacturing (AM) process, which combines AM and polymeric poling process, has been introduced to PVDF-based actuators. 160 The maximum strain of contraction perpendicular to the electric field can be 7% in low-frequency driving fields. 68 Therefore, RFPs have been applied in millimetre-sized legged robots and fish robots. 62,63 Except for the low strain values, the RFP actuation parameters transcend those of biological muscle (Figure 7(b)), with an output stress of 20 MPa during contraction and 45 MPa during the isometric state. 66,116,117 Furthermore, the fieldinduced deformation in high-elastic-modulus RFP can be operated at high frequency, which yields a strain rate of 2000%/s, a high work density of 1 MJ/m 3 , a high specific power of 4 kW/kg and a maximum efficiency of 80%. 66,118,119,161 The configuration of the RFP actuator is designed to be a single layer or multi-layer structure as the strain output at high frequency is smaller than that at low frequency.
For RFP actuators, the difficulty stems from the high level of driving field, which is close to the dielectric breakdown point. 155 The low sensitivity to the external stimulus limits the performance of the actuator and a large portion of the input energy will be extracted to match the load rather than act as an output force, which reduces the efficiency of the actuation system. 158 In addition, the ferroelectric-paraelectric transition occurs over a narrow range and involves a large hysteresis, which prevents an enhanced response from practical and general applications.

i-EAPs
CPs. CPs based on PPy are outstanding in this class due to their capability of producing more significant strain and stress, which are commonly employed as actuators. The oxidation/reduction reactions, which induce anionic dopants in PPy exchange for charge balance can produce electrochemomechanical deformation. The PPy-based actuator is fabricated using a twostep method. First, the substrate was coated with a PPy layer, which acts as the seed layer in the electrochemical deposition step. Second, the electrochemical deposition was accomplished by polymerizing the pyrrole monomer via electrochemical oxidation. 162 The strain produced by CPs is 2%-10%, and the maximum stress that is exerted is 34 MPa, even when actuated by a low voltage of approximately 2 V. [163][164][165] The PPy actuator devices can transfer electrical energy to macroscopic mechanical movements and mechanical work via the electrochemical reactions in polymer films. A bilayer or bimorph structure is employed to construct a bending PPy actuator, which includes a PPy layer and a passive layer. Macroscopic bending is caused by electrochemically driven length variations in the PPy film, which produces a stress gradient across the interface of the two films. In addition, the linear actuator is usually formed by a freestanding PPy film or yarns. To produce larger volume and length changes, other linear actuator configurations, such as multi-layered laminated films, folded films, bundles of films and roll-up structure, exist. [166][167][168][169] A work density of 100 kJ/m 3 and a specific power of 150 W/kg can be realised in PPybased CPs due to their high tensile strength and high stiffness. 122,124 These characteristics render CPs suitable as actuators in grippers and robots. [72][73][74][75] The selection of counterions and the design of structure have been two main topics: while electrochemomechanical deformation is limited by the counterions dopant, the structure of the actuator determines the robustness and actuation efficiency. However, the output strain of PPy actuators is small as small dopants are employed as counterions, even though the enhanced output strain has been gained by replacing these small counterions with larger counterions. A bilayer PPy actuator requires liquid electrolyte media and a metallic electrode to allow current flow, which increases the complexity and weight of the actuator. A major fraction of input electrical energy is wasted to produce the solvent discharge, which causes a low actuation efficiency. Moreover, these reactions cause pH variations and new chemicals, which promote the progressive deterioration of the actuation film and reduce the robustness of the actuator.
CNTs. CNT actuators are mainly powered electrochemically, where an electrical field causes swelling or contraction of the CNT when the mobile ions of solvent enter or leave the polymer. CNT actuators are explored in the form of sheets or films that act as bending actuators such as buckypaper actuators. 77,86 The CNT is mainly prepared by the arc-discharged method, where the CNT is synthesised on the electrode when running a DC current of 200 A at 20 V. 170 Other methods, such as the laser ablation method, chemical vapour deposition method, vapour-phase growth method and flame synthesis method, are employed to fabricate the CNTs. 147,[171][172][173] These actuators demonstrate relatively small strains between 0.1% and 3%. 86,126,174 A process for continuous production of CNT yarn fabrication was introduced, which enables the CNT yarn to be drawn from a multi-wall CNT forest. Compared with the form of sheet or film, the CNT yarn could be twisted, bent and knotted without structural failure.
Tensile CNT actuators could be constructed by twisted spun yarns or coiled yarns, and the output strains could be improved. 130,175 However, torsional CNT actuators could be formed by twist spun yarns, with two ends tethered. 176,177 Typically, the stress output of singlewalled CNTs (SWNTs) is approximately 0.75 MPa, and the maximum observed isometric actuator stress of SWNT actuators is currently 26 MPa. 128 Due to their high elastic moduli and giant tensile strength, CNTs have achieved a work density of 1 MJ/m 3 and a specific power of 270 W/kg. 127,128 These twisted and coiled yarns could be further processed into textiles by weaving and braiding, which could provide a larger tensile contraction and output force. 178 The CNT yarns of these actuators could also be driven electrothermally. Tensile or torsional actuation is achieved by a volume change of the yarn. To accommodate volume changes, guest materials, such as lie wax, are introduced to CNTs and transform a large guest expansion to a yarn expansion.
However, an electrolyte or guest material adds more volume and weight to the actuator system, which reduces the work density and specific power values. Moreover, the small strains of CNTs require some mechanical amplification. For twisted or coiled CNT actuators, small or even no space exists between neighbouring turns, and pre-strain mechanisms are essential due to the lack of space for an actuator to contract. These structures further increase the volume or weight of actuators. In addition, due to poor electromechanical coupling, CNT actuators require substantial energy recovery to achieve reasonable efficiencies. The bandwidths of these actuators are low due to the ion exchanging speed, which limits the strain rate. 179 Electrochemical driven CNT actuators rely on electrolytes to exchange ions, even though in some applications the liquid electrolytes are substituted. For example, a textile CNT actuator was driven with an ionically conducting gel instead of liquid electrolyte.
IPMCs. Manufacturing an IPMC actuator starts with the core part, an ionic polymer membrane and two conductive electrodes on both sides. Generally, IPMC manufacturing is divided into four steps: first, pretreatment (surface roughening); second, the initial compositing process, including impregnation-reduction and reductant permeation; third, the surface electroding process, including physical deposition and electroplating; and fourth, the ion exchange. 180 The major role of ionic polymers is to store the ions in the ion sites and maintain the required mobility for counterions across the membrane. Furthermore, ionic polymers provide the mechanical stress and force required for the soft IPMC actuators. While numerous polymers have been synthesised to achieve high performance, Nafion holds a place of prominence in the field due to its high proton conductivity and excellent mechanical properties. Nafion can be easily formed in any desired shape by casting or evaporating methods. 181,182 IPMCs based on Nafion can generate a large strain to 40%, with a stress of 3 MPa, under a voltage of less than 7 V due to their mechanical response. This large strain enables close emulation of natural muscle. IPMC actuators could use water as a solvent when DC voltage is applied. The success of these actuators in aqueous environments is a major advantage for the extensive use of IPMC actuators in various underwater robots. 183,184 IPMCs also have self-contained encapsulated electrolytes, which means that they can operate in dry conditions. Commonly, the IPMC actuator is developed using a thin Nafion film. Two polymer films are bonded to the IPMC actuator in a layered manner. Thus, bending is the basic motion of IPMCs, which is usually employed to form the cantilever actuator. Some linear actuators are developed based on IPMC actuators, where many basic units are connected in parallel or series, and utilise the buckles of IPMCs to achieve linear motion. 185 Therefore, IPMCs are extensively applied in robotic applications, such as micro-grippers and even microbiped and small quadruped robots. 132,186 The work density is approximately 5.5 kJ/m 3 , while the specific power is typically 2.56 W/kg. 187 When the IPMC actuator is applied in water, a corresponding increase in deformation beyond the initial state exists after recovering from the actuation stimulus. This phenomenon is referred to as the relaxation effect of IPMC, which causes serious instability of actuators. Although the use of ionic liquids instead of water as the solvent can relieve relaxation, the respond speed will decrease. Controlling the water content could eliminate relaxation but a slight variation in the water content will cause great difference in performance. Furthermore, due to the thickness of Nafion, which is nearly 180 mm, the output force is low. Apparently, an effective way to improve the output force is to fabricate IPMC with a thicker base membrane. Two methods are available to increase the thickness of Nafion film: multi-layer integration by hot pressing and solution casting. 181,182 However, the multi-layers are easily separated by long-term use, while the casting method will decrease the output strain and response rate.
As shown in Figure 7(c), the strain, stress, work density and specific power parameters are similar to those of biological muscle, and some of them even have values that are higher than the values for natural muscle. However, the strain rates of i-EAPs are less than 20%/s, even with a high applied voltage or closely spaced electrodes. 70,124 These slow actuation rates are produced by the slow transport of ions in electrolytes. Furthermore, the actuation efficiencies of i-EAPs are less than 1.5% due to their low electromechanical coupling. 77,88 Newly developed actuators. The newly developed actuators combine the strengths of different technologies to address many of the issues in actuators. As shown in Figure 7(d), these actuators achieve qualified actuation performances using different principles to linearly contract applications in a biological muscle-like fashion.
TFA. Twisted yarns of nano/microfibers, such as nylon, polyethylene, graphene and silk, can convert the volumetric expansion form to linear or torsional actuation. Among these twisted fibres, the inexpensive nylon polymer and fibre polymer chains in non-crystalline regions are less and have anisotropic thermal expansion behaviours, which enable relatively large linear actuation (to 49%) when heated. A single twisted nylon actuator is able to withstand loads that are more than 100 times heavier than human muscle with the same length and weight, which causes output stress of 22 MPa and specific work of 5.3 kW/kg. The maximum mechanical work during contraction is 2.17 kJ/m 3 . The bending and multi-directional motion can be achieved in twisted artificial muscle, which is fabricated via roller-pressing highly oriented nylon filament into rectangular or square cross-sectional beams. 92 Furthermore, textile construction is extensively employed as an amplifier of twisted actuators, where weaving and knitting are the two main textile processing methods. 188 Weaving has two perpendicular and individual tread systems -warp threads and weft threads -which could scale up force by parallel assembly of single twisted yarns. Knitting keeps the yarns together by loops that give latent potential for being easily deformable, which amplify the output strain by stretchable patterns. Textile processing permits the twisted yarn to build macroscopic actuation units and enables the potential to design more advanced architecture.
The challenge in developing a high-performance actuator is that actuation rates are limited by thermal transport times to access yarn volume. Usually, an electrical heating element involves helically wrapping the twisted muscle for fast heating but passive cooling still constrains the actuation rate. Even though a passive cooling method is immersed in the twisted muscle in water, which enables electrothermal actuation at 5 Hz to produce ;10% stroke, this method could only be employed in special environments. 94 Second, the maximum energy conversion efficiency during contraction is 1.08% for nylon twisted fibre, which needs to be improved. 94 HGA. The maximum swelling ratio of some hydrogel materials in deionised water is 8; 189 however, these materials have isotropic structures, and only homogeneous deformation can be achieved with uniform stimuli. For the actuator constructed with hydrogel, complex movements are essential. The preparation of internal anisotropic hydrogels has become more important for achieving complex deformation. Four main configurations, including the bilayer structure, gradient structure, patterned structure and oriented structure, are employed to build anisotropic actuators. 134 The maximum output strain of an actuator could be 20%, 135 and an output force of 1.8 MPa 136 can be achieved with a work density of 460 kJ/m 3 . By utilisation of an adjustable electric stimulation source, HGAs can achieve a highly controllable actuation state and tuneable actuation degree. Benefitting from these merits, these materials have potential for application in artificial muscles.
However, most HGAs depend on an aqueous environment, which creates two main challenges. First, the challenge for an ionic hydrogel is the leakage of solvent, which greatly affects the durability and stability of the actuator. Second, due to the slow water diffusioninduced hydrogel volume transition, the actuation process of a large-scale actuator is too slow to gain a real-time response, which caused a low strain rate (2%/s), low output power (2 3 10 -4 W) and low efficiency (0.164%). 137 HASEL. HASEL actuators have large output strains as fluidic actuators and can contract or elongate linearly without relying on aided structures. For a HASEL single-unit actuator, the maximum output strain is 170%, and a stress of approximately 0.3 MPa can be achieved when operating under a load of 1.5 kg. 100 Furthermore, the Maxwell pressure exerted in a HASEL actuator is independent of the electrode area, and the actuation strain and output force can be scaled by adjusting the ratio of the electrode area to the total area of an elastomeric shell or pouch. The output strain and force could be designed according to the requirement based on various HASEL configurations. In contrast to soft fluidic actuators, where inefficiencies and losses arise from fluid transport through systems of channels, and the additional weight of the mechanisms for producing pressurised fluids, the HASEL actuators generate hydraulic pressure locally via electrostatic forces and pump the liquid in reduced distances, which yields a fast actuation rate to 2176%/s and high conversion efficiency of 21% (similar to those achieved by DEA actuators). 138 Due to the great reduction in the design complexity, system weight and viscous loss, a single-unit HASEL actuator provides peak specific power of 586 W/kg and generates work of 64.4 kJ/m 3 during contraction. 138 However, the employment of liquid dielectrics in HASEL actuators increases the response speed and achieves high positional control. The thick elastomer shells (.1 mm) require high voltages to reach electric fields large enough for actuation, which have magnitudes in unit kV. In addition, the relatively large spaces for electrodes will be a drawback for certain applications.
Summary. The comprehensive comparison results are summarised as follows: 1. PAMs, such as the McKibben artificial muscle, are similar to biological muscle, and the actuation parameters are superior to those of biological muscle, except for a slightly smaller strain. However, PAMs are limited by the hysteresis and heavy gas tank in practice. 2. SMAs, such as the NiTi type, have large advantages in terms of the stress, work density and specific power, which originate from their high elastic moduli. For the same reason, SMAs exhibit low strain. The relatively low strain rate and low efficiency are caused by the thermal actuation method. 3. e-EAPs exhibit superiorities in strain rate and efficiency, high work density and specific power. The strain that is exerted depends on the elastic moduli of the materials. DEA materials with low elastic moduli have higher strain, while RFP materials with high elastic moduli have lower strain. However, the electric field required for actuation is high. 4. i-EAPs are actuated by a much lower voltage than e-EAPs and offer performances that are similar to those of biological muscle. However, the strain rate is low due to the slow transport of ions in electrolytes, and the efficiencies of i-EAPs are low due to the low electromechanical coupling. 5. Newly developed actuators display excellent biocompatibility, especially in the HGA, which could be applied to delivering medicine. Newly developed actuators also have excellent scalability, which could amplify their output strain and force in macro-scale actuators. However, the values of actuation parameters are relatively low compared with the other actuator technologies.
Artificial actuators have advantages over biological muscle in terms of the actuation parameters but challenges always exist when these artificial actuator technologies are applied in practice. Thus, artificial muscles cannot replace or perform as biological muscle. Therefore, further exploration of actuators with high performances is needed.

Future directions
The future trends of artificial actuators can be expected to fall into three different categories: (1) development of new methods for improving existing technologies, (2) application of bio-inspired design to artificial actuators and (3) establishment of bio-hybrid systems.

New methods for improving existing technologies
In the design of actuators for robotics, some methods have been proposed to eliminate the negative impact of actuation challenges. For conventional PAM-based robots, a power source for actuation, typically a gas tank, is required; it is often tethered to the robot. For untethered actuation, combustion-powered gas sources, which can release high amounts of energy and can scale with the size of the robot, were proposed. For example, 'Octobot' has self-contained aqueous hydrogen peroxide (H 2 O 2 ), which acts as a combustion source to excite the interior pneumatics (Figure 8(a)). 190 For e-EAPs, the main drawback is their use of high voltage (.1 kV typically) due to the high electric field that is needed. Using thin layers or creating materials with high dielectric constants, the required high voltage can be reduced. For i-EAPs, the strain rates are small due to the internal resistance of the polymers and electrodes and the slow ion diffusion rates inside the polymer. By increasing the applied potential, reducing the actuator thickness, or closely positioning the electrodes, the resulting actuation rates can be substantially improved. 70,124,191

Bio-inspired design
Unlike conventional actuators, which are composed of separate modular systems for sensing, actuation and control, novel artificial actuators should perform as biological muscle, which is capable of synthesising these features into its material architecture. Another important property of biological muscle is its scalability, which allows it to actuate large-or small-scale biological mechanisms. Additional actuator specifications, such as dynamic responses and morphological compatibility, should be taken into account when designing a robot. Inspiration can be obtained from nature, and researchers can apply bio-inspired topology, geometry and material to artificial actuator design to achieve versatile functions and adaptive interactions.
Regarding natural mechanisms, in addition to the actuation function, the body also offers cognition and control functions to offload computation from the brain, which is also known as 'morphological computation'. This process allows embodied agents, such as muscle, to not only function as actuators but also perform the task of sensing the environment and processing the information without participation of the brain. 195 For this reason, future artificial actuators could be combined with morphological computation, in which mechanical circuits are integrated such that the entire actuator can work as a computational resource for both perception and action. This actuation strategy would make the actuators work with rather than against their dynamics and reduce the complexity of the control architecture and the amount of computation, which enables more robust, faster and more efficient artificial actuator designs. In addition, based on the self-perception property, the morphological computation approach enables artificial actuators to switch their behaviours according to the environmental requirements. Artificial actuators can vary the stiffness distribution to satisfy the kinematic and dynamic requirements of the task. For example, the 'Yokoi hand', which is a 13 DOF robotic hand built from elastic and deformable materials for adjustment of its morphology to task requirements, can grasp different objects with the same simple control scheme (Figure 8(b)). 192

Bio-hybrid systems
Artificial actuators can also be formed by coupling biocompatible synthetic materials with contractile muscle tissues to form a bio-hybrid system. 196 Bio-hybrid actuators feature the merits of biological muscle, which are capable of self-repair and self-assembly and can perform actuation the same as in vivo. 197,198 However, we can control these bio-hybrid systems by external stimuli. For example, a jellyfish robot can be developed by coupling a monolayer of anisotropic cardiomyocytes with a soft PDMS elastomer. When phototactically following a light cue, this robot exhibits natural kinematics and locomotion (Figure 8(c)). 193 Another biohybrid robot consists of an antagonistic pair of skeletal muscle tissue that is combined with a flexible hydrogel substrate, which enables large and long-term actuation of the muscle tissue, that is, ;90°rotation of a joint with almost 1-week lifetime (Figure 8(d)). 194 Thus, bio-hybrid actuators based on tissue engineering are promising candidates for enhancing robot performance.

Conclusion
This study has presented the main differences between biological muscle and conventional actuators in robots. These differences indicate that bio-inspired and biomimetic robots require more actuator options for reproducing natural movements. To identify actuators that can perform as biological muscle, this study has reviewed the artificial actuator technologies that are extensively employed in robotics. These actuators are also known as artificial muscles, whose actuation is based on pneumatic, thermal, photonic and electrical stimuli. No actuator can satisfy all the actuation requirements. The advantages and shortcomings of artificial actuators must be distinguished by comparing their actuation parameters, not only to provide guidelines for choosing appropriate actuator materials but also to provide targets that are aimed at improving these actuator technologies. According to the reviewed data, this study performed single parameter comparisons and comprehensive performance comparisons, which offer new insights into the way actuators were evaluated and selected in this study and indicate future possible directions for enhancing artificial actuator performance.

Author contributions
L.R. and Zh.Q. were involved in conceptualization; W.L. and Ky.W. were involved in methodology; W.L. and H.L. were involved in writing-original draft preparation; and Zh.Q., L.R. and Lq.R. were involved in funding acquisition.

Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.

Funding
The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: This research was funded by the project of National