Fluid‐Driven High‐Performance Bionic Artificial Muscle with Adjustable Muscle Architecture

High‐performance artificial muscle is always the pursuit of researchers for robotics. Herein, a bionic artificial muscle is reported called “ExoMuscle” mimicking the sarcomere in skeletal muscle with a bio‐inspired structure to contract “myofilaments” enabling the artificial muscle to mimic the architecture of muscle such as parallel, fusiform, convergent, and pennation and beyond the performance of skeletal muscle. The reported actuators excel in various aspects compared with skeletal muscle including actuation stress (0.41–0.9 MPa), strain (50%), optimal length, velocity‐independence output, power density (10.94 kW kg−1), and efficiency (69.11%). With its own adjustable pennation architecture, it achieves variable actuation stress up to 0.9 MPa meanwhile maintaining high efficiency. Furthermore, ExoMuscle highly conforms to the anatomical complexity of the human body to cooperate with skeletal muscles closely opening the door for bio‐robotics, especially wearable robots.


Introduction
High-performance artificial muscle is always the pursuit of researchers for robotics to replace conventional motor or hydraulic systems in an unconstrained scenario. With the development of materials and manufacturing technologies, various ways have been proposed, as shown in Table S1, Supporting Information. To compare the performance fairly, researchers consent to the following performance metrics: actuation stress, actuation strain, response speed, specific work, specific power, and efficiency. [1,2] The negative pressure-driven method such as FOAM [3][4][5][6][7][8] has flexible nature but is limited by upper driving pressure to achieve high energy density. Positive-driven method [9][10][11][12] tends to have higher performance but is limited by large diameter increases such as McKibben and PPAM [13][14][15] and suffers from the viscosity of material such as Cavatappi. [16] Electrostatic-driven methods such as HASEL [17,18] have excellent efficiency, however, current material development restricted its obtaining high power density. The last method is thermo-driven one, [19] which tends to have extremely high work density; however, the highefficiency heat exchange system is not yet developed restricting its efficiency and response speed.
Biology tissues with hierarchical structures provide inspiration for science and engineering. [20][21][22][23][24][25] Mammalian skeletal muscle is the typical one that has been comprehensively studied by physiologists and biologists and revealed its subtlety and decency. Although previous research has demonstrated efficacy of mimicking the behavior on the level of muscle fiber using McKibben muscle, [26] till now, bionics seems to ignore its fundamental structures such as cross-bridge behavior in the sarcomere. Here, we report a bionic artificial muscle, called ExoMuscle, that bears a surprising resemblance to the morphology, architecture, and behavior of skeletal muscle and outperforms it in the concerned performances. Through sarcomere-like structure, ExoMuscle shows adjustable actuation stress, wider optimal length, higher contraction ratio, velocity-independent output, excellent power density, and actuation efficiency, which have yet to be reported. Further, the bio-mimetic architecture of ExoMuscle such as parallel, fusiform, convergent, and pennation shows tremendous potential of ExoMuscle in bio-robotics and wearable robots.

Design and Development of ExoMuscle
The hierarchy of skeletal muscle organization is shown in Figure 1a. The design and development of ExoMuscle are based on the "sliding filament hypothesis" [27] built from the observation that the I-band of sarcomere shortens during muscle fiber contractions, whereas the A-band does not as shown in Figure 1b and this mechanism shows unique ascending-plateau-descending behavior [28] as shown in Figure 2a because of the different acting range of thick and thin myofilaments. The mechanism of ExoMuscle is based on the serially connected sarcomere-like structures as shown in Figure 1c, composed of "myofilaments" of the same thickness and CB-tube. When the CB-tube opens, the edges of the CB-tube drive one group of "myofilaments" together to realize the cross-bridge behavior of "myofilaments" just like what happens in A-band as shown in Video S1, Supporting Information. The CB-tube is named after the functions it acts, cross-bridge tube. The above sarcomere-like structures or single-module ExoMuscle composed of CB-tube and "myofilaments" (wires) are connected serially to form a whole artificial muscle, as shown in Figure S2 and S4, Supporting Information. The design concept of the CB-tube is based on the previous practices of researchers that although in the past decades some "bionics" actuators claim that they have the same behavior as muscles fascicles, [29,30] their force-length ratios only cover the ascending limbs of that of muscle fascicles as shown in Figure 2b. However, when compared with the muscle in vivo which runs around the optimal length or plateau region [31] (Figure 2a), the force-length relations are not satisfied. That is, when the actuators are deployed into bio-robotics, the output torque-angle relations cannot keep stable as most joints or causes insufficient working range. [32] The effect is much more obvious when the actuators built aligns with the joint to provide assistance because of the increased lever arm. The design concept and details of the CB-tube are shown in the corresponding Supporting Information and could extend the ratio of plateau region or optimal length to fulfill the demands on bio-robotics or the scenario when working in alignment with skeletal muscle. Based on the anisotropic design of the CB-tube, ExoMuscle could have much higher actuation stress, higher actuation strain, and wider optimal length than skeletal muscle and could work in alignment with skeletal muscle fitly. Hence, the reported artificial muscle is called Exo-Muscle. The ExoMuscle could adjust its architecture like skeletal muscles such as parallel, fusiform, convergent, and pennation as shown in Figure 1e-g to achieve variable shape and actuation stress to adapt to different scenarios. Once the bionic mechanism is demonstrated, the comparisons between ExoMuscle and skeletal muscle were tested around the aforementioned metrics. Other interesting nonquantifiable metrics shown in mammalian skeletal muscles are also tested such as output-velocity relationships and variation of muscle architectures.

Basic Comparison between ExoMuscle and Skeletal Muscle
We characterize the force--length relations through a series of quasi-static experiments. A single-module ExoMuscle, with the CB-tube cross-sectional area (CSA) of 9 cm 2 shown in Figure S1i, Supporting Information, is tested on the bench shown in Figure S5a, Supporting Information. The detailed result is shown in Figure S5e, Supporting Information. A blocked force of %360 N and contraction ratio of %50% was generated using 900 kPa positive pressure supply and shows a wider plateau region (output force is stable within an error of 10%) than that of active force of the skeletal muscle. It also demonstrates that ExoMuscle shows actuation stress of 410 kPa more than four times of 100 kPa of skeletal muscle (sustainable state) [1] and higher than 300 kPa at optimal fiber length during fully activated (although this state muscle could only contract two times with stored ATP [33] ). This stable output nature can be explained using a model based on the principle of virtual work, as shown in the Modeling section. It is observed that the proportion of the plateau region of the force-length plot reduces with pressure due to the extensible nature of the wires. When pressurized over 500 kPa, this phenomenon becomes more obvious, and the opening angle of the CB-tubes increases with pressure. Figure 2b shows a comparison among force-length relations of ExoMuscle, muscle fiber, muscle in vivo, and other actuators where the output force and length are normalized by maximum force P 0 and shortest length L s . The force-length relations of ExoMuscle are adopted from the single-module ExoMuscle at 400 kPa. And the result shows it covers the plateau region of the muscle fascicle and could have more stable output than other actuators.
Ten-module ExoMuscle are also characterized ( Figure S5b, Supporting Information) to study the effect of a series connection of sarcomere-like structure. The plateau region extends over 75% in all pressures, with fluctuations evident in the plot. This phenomenon is caused by the differences in the stiffness of the 3D-printed CB-tubes. Hence, they inflate at different rates rather than simultaneously, which extends the proportion of the plateau region or optimal length with some decreases in the output force.
The force-length relations describe ExoMuscle and skeletal muscle's behavior at a constant length, i.e., isometric conditions. However, natural muscle active force is highly dependent on its velocities, [34] showing an inverse proportional relation due to the finite amount of time for cross-bridge attach, as shown in Figure 2c. The output force of ExoMuscle keeps quasi constant with the velocity increasing to 4.5 times of fiber lengths per second (the velocity is also the maximum reciprocating velocity of the test bench), i.e., a velocity-independence behavior in the operating range of skeletal muscle. (The detailed result is shown in Figure S6, Supporting Information.) The unloaded maximum contraction velocity obtained is about 2.5 times of fiber length per second. [35] However, as we tested, the contraction velocity of ExoMuscle reaches 4.4 m s À1 at 1 kg load as shown in Figure 2g and S7b, Supporting Information, i.e., 22 times of fiber length per second much faster than skeletal muscle fiber. Meanwhile, the maximum specific work of ExoMuscle is The main power of muscle is from the sarcomere structure consisting of myofilaments in muscle fibers. This figure also shows the parallel structure of muscle where muscle fibers are parallel to tendons such as biceps brachii and sartorius. b) Sarcomere structure in muscle fibers. Thick and thin myofilaments of sarcomere cross with each other showing a cross-bridge region or A-band and thin filament region or I-band. They use the energy of ATP to contract the muscle fiber c) ExoMuscle contains several sarcomere-like structures connected serially, whereas the "myofilaments" have the same thickness and cross each other to contract by the expansion of the CB-tube same as what happens in A-band d) Three-dimensional diagram of ExoMuscle. When CB-tubes inflate, their edges pull "myofilaments" together to contract the whole muscle e) The unipennate muscle such as extensor digitorum shows angles between muscle fibers and tendon enabling much more muscle fibers in limited cross-sectional area to increase actuation stress. The bionic ExoMuscle also has its pennation way to increase actuation stress-like skeletal muscle, whereas its pennation is from the rotation of the CB-tube from "myofilaments." f ) Bipennate muscle such as rectus femoris is a variant of pennate muscle where the muscle fibers have two pennation angles. The bionic ExoMuscle could mimics this behavior to put two symmetric unipennate ones together. g) Convergent muscle such as pectoralis major is variant of parallel muscle, where muscle fibers convergent to a smaller area. The bionic ExoMuscle could mimic this behavior to put three parallel ones together and be partially pressurized as the selective activation of this type.
www.advancedsciencenews.com www.advintellsyst.com 4.34 kJ kg À1 tested at 15 kg load (0.04 kJ kg À1 of skeletal muscle [36] ) and the maximum power is 10.94 kW kg À1 obtained with 5 kg weight (0.28 kW kg À1 of skeletal muscle [36] ) as shown in Figure 2c,d indicating a load-dependence behavior. And the average velocity of ExoMuscle at 20 kg load reaches 5 relative length per second, still exceeding the non-load velocity of muscle fiber.
As a unique characterization of fluid-driven method, it is worth noting that the power and work could be adjusted with source pressure and flow rate. [37] The same load as the maximum one of test for specific power, the lifetime of ExoMuscle was tested. Figure S9c, Supporting Information, shows the lifetime test result showing no degradation in displacement or output force during the lifting process in the first 10 000 cycles, as the first seven cycles and last seven cycles all have the same overshoot as Figure S9c, Supporting Information, and no decrease is found. Due to the inconsistency of 3D printing, the lifetime varies with the quality of the 3D printing. In general, the lifetime of 3D-printed ExoMuscle is on the order of 10 4 cycles. It seems very hard to achieve the 10 9 cycles lifetime as a skeletal muscle because muscle contains fabrication machinery for its manufacture and regeneration. [38] However, with the cheap and fast manufacturing method shown in the Method section, simply ExoMuscle has a wider optimal length covering an operating region of muscles, whereas other actuators cannot cover this range. c) Force-velocity relations of muscle and ExoMuscle. Muscle fascicle shows a velocity-related output, whereas ExoMuscle does not. d) Specific work changes with load and reaches the highest value at 15 kg exceeding that of muscle much. The velocity of ExoMuscle decreases with load e) Specific peak power and average power of ExoMuscle changes with load and reaches the highest value at 5 kg much exceeding all that of muscles. f ) The efficiency of ExoMuscle increases with load and reaches a maximum value at 30 kg g) Radar graph for comparing between ExoMuscle and muscle fibers h) Three ExoMuscle with cross-sectional area (CSA) of 5 cm 2 forming a fusiform muscle drives a bucket with 20 kg weight up i) Three ExoMuscle with CSA of 5 cm 2 forming a convergent muscle drives the bucket up. Two of three ExoMuscles are pressurized lifting the bucket left j) and right k) up to mimic the selective activation of convergent muscle l) a parallel ExoMuscle with CSA of 2.5 cm 2 drives a 4 kg weight. The corresponding video is shown in Video S2, Supporting Information.
www.advancedsciencenews.com www.advintellsyst.com changing one of the tubes of ExoMuscle could largely extend its lifetime. And the damage is found on the horizontal line of 3D printing in the opening of the CB-tube. It is anticipated that replacing 3D printing with blow molding to obtain homogenized structures would greatly prolong the lifetime of ExoMuscle. One interesting phenomenon is that the chemo-mechanical efficiency of isolated muscle fiber of frog could reach larger than 40%, [39] whereas that of the whole muscle in vivo is only 15% for concentric contraction [40] showing similar velocity dependence behavior like that of output force. [41] This may result from the interactions between muscle fibers and extracellular matrix (ECM) [42,43] even the interactions between the bulging nature of muscles of agonist group. [44] It is then questionable whether it is feasible to obtain a highly efficient macroscopic robot through a bottom-up approach if there are interactions with energy storage or dissipative behavior such as friction. As shown in Figure 2f, the efficiency of ExoMuscle increases with load and could reach as high as 69.11% at 30 kg load in a top-down design concept, much higher than muscle fibers and the whole muscle.

Muscle Architectures of ExoMuscle and Skeletal Muscle
Although muscle fibers have aforementioned drawbacks such as relative low actuation stress and velocity-dependent force, human has evolved different muscle architectures to adapt this nature such as parallel-type and pennate-type architecture as shown in Figure 1, which have a profound impact on the muscle performance. [45,46] Muscle fibers of parallel type are aligned with tendons and have some variations such as fusiform and convergent ones. Parallel-type muscle maintains the performance of muscle fibers and the convergent one has its own advantage such as selective activation of muscle fibers to achieve multi-DOF movement. ExoMuscle could easily mimic the aforementioned structure. The original ExoMuscle is like the parallel one as shown in Figure 2a, and the fusiform one is formed when three or more ExoMuscles are put together as shown in Figure 2h. When one end of three ExoMuscles is put together and the other ends are separated and fixed, the convergent one is formed as shown in Figure 2i, and could partly be pressurized to mimic the selective activation of pectoralis major as shown in Figure 2j,k. To demonstrate the scalability of ExoMuscle, one with a miniature thickness (5 mm) is built and could drive 4 kg weight up, as shown in Figure 2l and S4, Supporting Information.
Muscle fibers of the pennate-type muscle are oblique to the direction of tendons compared with parallel ones as shown in Figure 3a, which enables muscle with higher force in the constrained CSA. The area vertical to the muscle fibers is called the physical cross-sectional area (PCSA), which is always used to estimate the maximum muscle force. [47] For example, pennation enables the soleus muscle to achieve 220 cm 2 PCSA in only 30 cm 2 CSA increasing the actuation stress 7 times, i.e., much enhancing the output force in the constrained area but with a lower contraction ratio. [48] And the pennation also brings various gear ratios of transmission making muscle fibers work in a highly efficient velocity region, [49,50] which is considered regulated by connective tissues and the bulk of muscle fiber during actuation. However, pennation is not always a high-efficiency mechanism, it seems a compromise of the overall efficiency because with the increase of pennation angle, more strain energy is stored at ECM, [51] but it truly provides an inspiration to increase actuation stress in a limited CSA. Although ExoMuscle does not contain the "ECM" between "myofilaments" (wires) to hold the shape of muscles, it also has its own way to achieve the same pennation behavior of muscle. Without "ECM" among wires, the pennation of ExoMuscle would result in rotation between CB-tube and wires, as shown in Figure 3d. Hence, ExoMuscle could achieve various pennation angles, as shown in Figure 3b,c. Tilting the CB-tubes makes more fibers contained in limited CSA. As shown in Figure 3e, with the increases of pennation angle to 70 ∘ , the actuation stress increases to 0.9 MPa and efficiency decreases to 51.14% because the friction between the wires as shown in the backside plot of Figure 3b,c; however, it is still a highly efficient way to increase actuation stress. The test method and detailed results are shown in Figure S5c,d,g, Supporting Information. Figure 3f shows the unipennate ExoMuscle with a 45 ∘ pennation angle, and Figure 3g shows the bipennate ExoMuscle with a 70 ∘ pennation angle.

Application Demonstrations
The compliant and flexible nature of ExoMuscle enables the force distributed on the two wires when the wires are staggered as shown in Figure 4f, which shows similar behavior to the differential mechanism used in the rigid prosthetic hand. [52][53][54][55] The prosthetic hand has compliant tendons in the fingers, which allows them to return to the initial position when the force is released. When there is no object in hand, the four fingers clench together during actuation, and when the tool is placed on the hand, the four fingers will adapt to the shape of the tool to clamp it with two fingers, and the other two fingers will flex to the maximum angles as shown in Figure 4f. When the hand clenches a smartphone, the fingers touch the edges of the phone one by one because of the compliance of ExoMuscle. This demonstrates the driving ability and differential nature of ExoMuscle. We also show a prospective application on bio-robotics in pectoralis major as shown in Figure 4a, which shows similar behavior of pectoralis major.
Because the motion of wrist extension is currently omitted in the robot-assist stroke rehabilitation training [56][57][58][59] leading to the unnatural motion of grasping which may affect the training efficacy, a parallel-type ExoMuscle with a CSA of 6 cm 2 has been deployed in a hand extension robot to replace the function of Extensor carpi ulnaris (ECU), extensor carpi radialis brevis (ECRB), and extensor carpi radialis longus (ECRL) for wrist extension as they function as agonist muscles for wrist extension, as shown in Figure 4b. One end of a six-module ExoMuscle is fixed on the back of the hand and the other end is fixed on the forearm. To check whether the force of ExoMuscle meets the demands of wrist extension, the comparison of the CSA of ExoMuscle and the sum of PCSA of ECU, ECRB, and ECRL is enough because the actuation stress of ExoMuscle and skeletal muscle is comparable. As previous research on the health and young subject, [60] the summation is about 7.7 cm 2 , that is, the ExoMuscle could replace those muscles for wrist extension as shown in Figure 3b completely and even lift weights. To prove the controllability meanwhile, We deployed www.advancedsciencenews.com www.advintellsyst.com www.advancedsciencenews.com www.advintellsyst.com a wire encoder to acquire the length of the ExoMuscle and a close-loop control to finish sinusoidal tracking of 1 kg loaded hand extension, as shown in Figure 4b. Because of the higher lever arm of the outer body conditions, the necessity of optimal length is obvious. With a larger portion of optimal length, ExoMuscle could drive the wrist in a range of 45 degrees effortlessly without the participation of according muscles. Future work will detail the design to add artificial ligaments [24] and use more modules to retrieve natural full range of motion for stroke patient.
In Figure 4d, we also show a prospective application of ExoMuscle with the assistance of plantarflexion with a CSA of 10 cm 2 . As shown in the figure, ExoMuscle connects with the foot through a strap surrounding ankle, hence the lever arm of ExoMuscle is similar to that of the soleus (SOL) and gastrocnemeus (GA). ExoMuscle can cooperate with the gastrocnemius and soleus closely. The corresponding data are provided in Figure 4d; gastrocnemius muscle activation decreases by 19.1% and soleus muscle activation decreases by 29.9% at a speed of 3 km h À1 . The control method in this scenario is the Figure 4. Application demonstrations. ExoMuscle covers motions of major joints, showing tremendous potential in bio-robotics and wearable robotics. a) Three ExoMuscle forming a convergent muscle to mimic the behavior of pectoralis major on a skeleton model. b) Because the actuation stress of ExoMuscle is comparable to skeletal muscle, when close fitly use, comparing the physical cross sectional area (PCSA) of muscles is enough for designing. ExoMuscle drives the wrist to lift a 1 kg weight in a sinusoidal wave, demonstrating its driving ability and controllability. c) One end of ExoMuscle is connected on shoulder, and the other end is fixed on forearm to assist elbow flexion. The corresponding test results are shown in Figure S10a, Supporting Information, d) Parallel ExoMuscle with 10 cm 2 CSA could work with soleus and gastrocnemius closely to assist plantarflexion. Parallel ExoMuscle reduces SOL and GA muscle activation and would reduce more with pennate ones. e) Two ExoMuscles converge on arm to assist shoulder flexion and abduction. f ) The differential nature of ExoMuscle enables driving four fingers with only one parallel ExoMuscle. The figure shows the differential process when the hand holds a smartphone. And ExoMuscle drives a prosthetic hand to clench a tool with two fingers. g) Two 45 ∘ pennate ExoMuscles assist hip flexion and one 70 ∘ pennate ExoMuscle assists knee extension. The corresponding video is shown in Video S4, Supporting Information.
www.advancedsciencenews.com www.advintellsyst.com same as that used in ref. [61]. It is anticipated that with the pennation of ExoMuscle, the CSA could increases to 40 cm 2 and accounts for more part of joint torques, [48] and the result will be done in future research considering the complex internal energy recycles [49,62] with more complex control method. To demonstrate the tremendous potential in bio-robotics and wearable robots, we also show the prospective assisting application on elbow flexion (Figure 4c) where ExoMuscle reduces 37.66% muscle activation of biceps brachii when the user lifts a 4 kg weight as shown in Figure S10a, Supporting Information, shoulder abduction and flexion (Figure 4e) where ExoMuscle reduces 35.73% muscle activation of middle deltoid when shoulder abducts as shown in Figure S10b, Supporting Information, hip flexion and knee extension (Figure 4g), which covers all major joints of human motion.

Discussion
In this article, we introduce a bionic artificial muscle, ExoMuscle, which mimics the behavior of cross-bridge in the sarcomere of skeletal muscles. With biological inspired CB-tube and "myofilaments," the ExoMuscle demonstrates performance superior to skeletal muscle and has adjustable muscle architecture such as parallel, fusiform, convergent, unipennate, and bipennate to further increase its flexibility and compatibility.
Here are some discussions about scaling, clustering, system optimization, hysteresis, and practical use compared with other arts.
As we tested, if the diameter of CB-tube increases and the wall thickness increases by the same multiple, the actuation stress and efficiency could be considered as the same because the strain energy increases with input energy at same scale factor. This is a cost-effective way to increase output force at same pressure although increasing the demand of input flow. For example, if the fluid source is with low pressure and high flow rate, increasing the diameter of CB-tube is suitable for the system configuration and vice versa. Because the CB-tube is always kept facing outside of the wires, the motion of users is not affected by increasing of diameter when close-fitting. Compared with scaling up, scaling down involves the problem of fabrication. Because of the limitation of minimum printing width of 3 d printer, the wall thickness of CB-tube could not be decreased by same multiple as diameter. Hence to prevent wall broken, the wall thickness of CBtube with lower diameter should be increased to compromise the limitation. However, it could be anticipated that using blowmolding method, this limitation of 3D-printing could be avoided.
If the CB-tube is lengthened, the actuation stress slightly increases as the CB-tube could be divided into three parts including two margin portions and one middle portion. Because the expansion of the margin portion is restricted by bottom surface, the actuation stress is affected, whereas the middle portion is not. Lengthening CB-tube means increasing the portion of middle one. Hence, the actuation stress increases. The same is for efficiency. It is noteworthy that the length to diameter ratio should not be too low to prevent the above effect. For power density, it is affected by input pressure difference, length, and diameter of air pipe, [37] we will not discuss this part in detail. Contraction ratio keeps constant during the process of scaling as modeled.
As for clustering, because the relative thickness increase during actuation is lower than PPAM, [13,14] Cavatappi [63] and Peano muscle, [64,65] and the dense wires forms a smooth division for tubes and wires, when actuators are parallelized and stacked together, the overall performance is only affected by the friction between tubes of one actuator and wires of another actuator. The motion is not affected by stuck caused by bulge among actuators as well, which is the advantage of this design compared with others, as shown in Figure 2h-k and Video S2, Supporting Information. However, this configuration is not recommended to obtain large sectional area because compared with increasing diameter, length, or CSA of CB-tube to obtain larger force at same pressure, stacking brings unpredictable friction to control, unless the overlapping area is low as shown in Figure 4c for multi-DOF shoulder assistance to minimize this effect.
As for optimization of the performance, this is a complicated engineering problem, as large number of FEM trials and manufacturing are required considering different materials. The dimension shown in the paper is the stable parameters for the material. But if we only foresee the upper bound of the performance, we claim that the performance could be continuously improved if a material with low flexural modulus and large tensile modulus is achieved in future. As we modeled, lower flexural modulus or bending modulus means the energy stored in the material is less, and larger tensile modulus means the actuator could bears higher pressure. The above two ways seem not conflicting each other as the filament is an example. This is also the advantage of this structure using bending of material to create motion compared with other structure using stretching of material, where increasing tensile strength to bear larger pressure causes decreasing of contraction ratio or output force at same pressure which is a paradox.
As for hysteresis, this phenomenon is universal and happens in various fields. [66] The hysteresis loss is also the focus in practical use of an actuator. The isobaric force-displacement relation when single-module and multi-module parallel ExoMuscle repeat three cycles of contraction and stretching is obtained as shown in Figure S12, Supporting Information.
It could be seen that the hysteresis is obvious in the process, showing a unique noncentrosymmetric manner. The result shows about 40.7% hysteresis loss for one module. With the increasing of pressure, there appears a local extreme point at the limiting ascending branch of single-module one. At the pressure where single module appears the extreme point, multimodule ExoMuscle has fluctuations on corresponding part, and meanwhile the flat part of output force of multiple one is extended compared with single-module one. We hypothesize that this phenomenon is related to the process of above hysteresis that improves its effective driven portion. Because the output characteristics among CB-tubes are various, hence when the pressure increases, the CB-tube with the largest output force first expands at its limiting ascending branch of Figure S12a, Supporting Information, and other tubes are at their descending branch. When the output force of the first tube decreases, the tube with second large output force starts to expand from its descending branch to ascending branch. The process was repeated as all the tubes are inflated, which forms fluctuations on limiting ascending limb of Figure S12b, Supporting Information, one by one and extends the effective driving www.advancedsciencenews.com www.advintellsyst.com portion. This effect is very useful in scenarios where a large CSA muscle is connected in series with multiple muscles with small CSA in parallel to generate multi-DOF motion and large output forces simultaneously. When the multimodule muscle contracts because of the series connection, the lower bound of contraction force is dominated by the modules with lower output force, whereas when the muscle is extended passively, the upper bound of output force is determined by the modules with higher forces. During the extension process, the modules with higher forces does not return to the zero-opening angle, and the opening angle reduces less compared with tubes with lower forces, whereas the opening angle of the modules with lower forces will be over closed, introducing hysteresis loss of the viscoelasticity material. This increased force during extension is fast released at the beginning of the contraction process as shown in Figure S12, Supporting Information, which is a different hysteresis loop compared with that of a typical artificial muscle. This phenomenon is that at the same length and pressure, the opening angles of modules at contraction and extension process are totally different, i.e., the muscle is extended to a totally different state using large energy although length and pressure are the same. This is why the hysteresis loss of the multimodule muscle is larger than single module one. As for single-module muscle, the largest viscoelasticity loss is around the maximum contraction ratio. Around this state, according to the observation, the two concave faces of the tube are further stretched and when the muscle is passively extended, there requires more force to restore its previous strain state. This means that the energy efficient way is to prevent the fully expansion. It is worth noting that the wires of the muscle only limit the perimeter of the tubes with concave section, which is different to McKibben-like muscle where the wires fully determine the shape of the tube with convex section. Hence, the detailed research will be conducted in future to determine whether the hysteresis source is like the increased part of multimodule one where at same pressure and length the states are totally different, requiring more energy to extend it to previous length.
Although hysteresis brings difficulty in control, this process is too universal and have similar mathematical representation such as Preisach model, [66][67][68] Prandtl-Ishlinskii model, [69,70] etc. in different physical process. As the next step for precise control, we have tried Preisach model to compensate this process and achieve reasonable results because as we tested the hysteresis for the single module one shows no accommodation and does not change with velocity where viscosity effect is not too obvious. By compensating this phenomenon using inverse solution of this model, the control accuracy is expected to increase. Then we will deduce the multimodule model based on the accurate model of single module and above hypothesis.
The data collected in Table S1, Supporting Information, can be used for comparison with other artificial muscles ( Figure S11, Supporting Information). ExoMuscle clearly exhibits a more excellent performance than the other artificial muscles. Note that all the actuators chosen are the most classic ones in their type, and ExoMuscle provides the highest specific work, efficiency, and much larger actuation stress than nonthermo-driven type meanwhile fixing the slow response of thermo-driven ones. Given its excellent performance and variable architectures, we conclude that ExoMuscle is a suitable option for future bio-robots, especially wearable robots for all major joints drive and assistance as shown in Figure 4.
Here is some benefit brought by the metric improving in actual use. For example, in the scenario on wrist extension in paper, the actual peak pressure is around 300-600 kPa only. And for this level of pressure, commercial miniature pump fully meets the demand. The commercial McKibben muscle [71] usually works around 500-800 kPa (10-25% contraction ratio), whereas the Cavatappi [63] usually works around 100-200 psi (700-1400 kPa). That means these two types require large pressure to offset their own elasticity, which is a sign of low efficiency at low pressures. For the Peano muscle, [64,65] the actuation pressure is low; however, it only makes the no-load motion possible [72] or even worse the low pressure usually does not meet the demand in many assisting scenarios as we tested. Compared with the three types, ExoMuscle improves the usability at low pressures by increasing efficiency meanwhile achieves great performances at high pressures in a decent manner to let the exoskeleton compact, flexible, and close fitting, which releases the strict demand on pump and push the limit to a wide range of use from light to heavy load.

Experimental Section
The CB-tubes in this paper were 3D-printed using thermoplastic polyurethane material with shore hardness 83 A (EFLEX, ESUN cooperation). For stable manufacturing quality, the CB-tube was 3D-printed with a preopening angle of 40 ∘ and then annealed at 130°C for 30 min before cooling down under room temperature to keep the stretching state after removal of external forces. Each tube was wound by thin filaments (0.05 mm) during inflation and the filaments were glued on the side of the CB-tubes ( Figure S3c, Supporting Information). Wires were wound densely around the tubes clockwise and counterclockwise and then glued to the wall of the tubes. Finally, wires were sewn on straps using a sewing machine. All the experimental methods used in the main text had detailed corresponding part in Supporting Information and marked in the corresponding place.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.