# Artificial muscle

> Source: https://aiwiki.ai/wiki/artificial_muscle
> Updated: 2026-07-14
> Categories: AI Hardware, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Artificial muscle** is an umbrella term for actuator technologies that generate motion by contracting along a line, the way biological muscle shortens to pull a tendon, instead of by spinning a shaft the way a conventional motor does. The category covers several unrelated mechanisms, pressurized gas, pressurized liquid, electrostatic force, heat-driven metal phase change, and thermally driven polymer fiber, that share only this linear output. In humanoid robotics the term has taken on renewed relevance since the mid-2020s as companies such as [Clone Robotics](/wiki/clone_robotics) and [Sanctuary AI](/wiki/sanctuary_ai) have built dexterous hands and android-style bodies around fluid-driven muscle fibers instead of the [actuator](/wiki/actuator) stacks, a rotary motor paired with a [harmonic drive](/wiki/harmonic_drive) or ball screw, that dominate most of the industry.

**In brief:** an artificial muscle is any actuator that pulls rather than turns. Instead of a motor spinning a gear train, something inflates a tube, charges a pair of plates, heats a wire, or untwists a coiled fiber, and the whole element gets shorter, closer to how a biceps works than how a servo motor works. No single technology matches biological muscle on every axis; each family trades away strain, speed, efficiency, or simplicity to win on the others.

## How artificial muscles work

Every artificial-muscle family converts an input, gas pressure, liquid pressure, an electric field, or heat, directly into a change in the length of a soft or fibrous element anchored at both ends, similar to how a [tendon-driven](/wiki/tendon_driven) system routes force from a remote motor. This differs from most robot joints today, where a rotary motor spins quickly with low torque and a reducer (a harmonic drive, cycloidal drive, or ball screw) trades that speed for torque. Artificial muscles skip the reducer stage because contraction is itself the useful motion.

That simplicity is also the category's central problem. A conventional electric actuator is a mature, modular stack of motor, bearing, and gearbox available off a catalog. An artificial muscle usually has to be sized, sealed, and controlled as a bespoke element, and most families below are less efficient or harder to control precisely than an equivalent motor-driven joint, which has kept them a minority approach in humanoids even as they remained a steady research topic[1][2].

## Types and variants

### Pneumatic artificial muscles (McKibben / PAM)

The oldest and most common family is the pneumatic artificial muscle (PAM), commonly called a McKibben muscle after Joseph L. McKibben, a Los Alamos National Laboratory physicist who built the first version in 1957 as an orthotic aid for his daughter, whose hands had been paralyzed by polio[3]. A McKibben muscle is an elastomer bladder wrapped in a braided sleeve; compressed air inflates the bladder, and the braid geometry converts radial expansion into axial contraction, similar to how a Chinese finger trap shortens when pulled. Braids under about 54.7 degrees contract on inflation; braids above that angle extend instead[4]. Bridgestone commercialized the design in Japan in the 1980s as the Rubbertuator, and it remains in production today, most visibly as [Festo](/wiki/festo)'s Fluidic Muscle[3].

Festo, a family-owned German industrial-automation company founded in 1925, sells its Fluidic Muscle DMSP as a catalog component in bore sizes from about 5 to 40 millimeters, capable of hundreds of newtons of force (630 N for a 10-millimeter bore) and cycle rates up to roughly 150 Hz, with best repeatability below about 9 percent contraction rather than at maximum stroke[5][6]. Festo's separate Bionic Learning Network has also built pneumatic demonstrators, including the elephant-trunk-inspired BionicSoftArm and BionicMotionRobot (12 degrees of freedom, roughly 3 kg payload) and the ExoHand exoskeleton glove, research platforms rather than shipping products[7][8].

PAMs are compliant and simple, but compressed air is compressible, so the muscle behaves like a nonlinear spring that is hard to hold at a precise position without extra sensing. That tradeoff drove an early adopter, Shadow Robot Company, away from pneumatics: its air-muscle hand, introduced around 2004 with 24 degrees of freedom driven by rubber-tube-in-mesh muscles pulling tendons, was the company's flagship product until about 2010, when the [Shadow Hand](/wiki/shadow_robot) switched to internal electric motors for better precision, a configuration its modern hand still uses[9].

### Hydraulic artificial muscles

Replacing air with an incompressible liquid in the same braided-sleeve geometry produces a hydraulic artificial muscle. Because liquid barely compresses, it responds faster, holds position more precisely, and wastes less energy than the pneumatic version of the same design: a 2021 study measured roughly 60 percent quasi-static efficiency for a latex-bladder muscle driven hydraulically, versus roughly 25 percent driven pneumatically[10]. In 2017, Bridgestone and the Tokyo Institute of Technology demonstrated a hydraulic McKibben-type muscle rated for 5 megapascals of internal pressure, 15 millimeters in outer diameter, about 7 kilonewtons of force at up to 30 percent contraction, and a claimed strength-to-weight ratio five to ten times a comparable electric motor or conventional hydraulic cylinder[11][12].

The tradeoff is plumbing: a pump, reservoir, and valves that must be fast and essentially leak-free, easy to meet in an excavator but historically hard inside a human-sized limb. Sanctuary AI, a Vancouver, Canada, humanoid company founded in 2018, miniaturized the valves themselves: its Phoenix hand uses a 21-degree-of-freedom hydraulic design built around coin-sized valves the company says are roughly 50 times faster and six times cheaper than off-the-shelf equivalents, cycle-tested past two billion actuations without leakage[13][14]. Clone Robotics, a Poland-based startup founded in 2021, took a related path with its "Myofiber" muscles, McKibben-derived fibers the company says contract more than 30 percent unloaded, respond in under 50 milliseconds, and generate about a kilogram of force from a three-gram fiber[15]. Clone's early Protoclone and Torso demonstrators ran on compressed air, with water routed through the frame mainly to cool the muscles, and the company itself described that generation as "pneumatic actuation with off-the-shelf valves"[16][17]. Its newer Alpha android, unveiled in 2026, is described by the company and independent coverage as running on hydraulic Myofiber muscles instead, with a roughly 500-watt "hydraulic heart" pump and low-power custom "Aquajet" valves replacing the earlier pneumatic ones[16]. Public demonstrations of both companies' hydraulic systems have so far shown supported or bench-mounted motion rather than untethered walking, so a self-contained walking android remains a roadmap item[13][17].

### Electrostatic and electrohydraulic actuators

A dielectric elastomer actuator (DEA) is a thin sheet of soft, insulating polymer sandwiched between two compliant electrodes. A high voltage across the sheet squeezes it electrostatically, making the film thinner and wider, and if the geometry channels that spreading into a pull on a load, the device contracts like a muscle. Ron Pelrine and colleagues at SRI International began this work in the early 1990s as the electroactive polymer artificial muscle; SRI spun it out as Artificial Muscle, Inc. in 2003, won a Frost & Sullivan innovation award in 2006, and was acquired by Bayer MaterialScience in 2010, which commercialized a haptic-feedback line under the ViviTouch brand[18]. DEAs can be fast, up to roughly 1 kilohertz in some designs, and efficient, since electrostatic actuation wastes comparatively little energy as heat; under large prestrain, lab demonstrations have reported area strains exceeding 300 percent, though practical actuators run at a fraction of that[19]. The recurring obstacle is voltage: classic DEAs need several kilovolts to drive, awkward inside a battery-powered robot, and lowering that toward tens or a few hundred volts is an active research area[20].

A related but distinct approach, hydraulically amplified self-healing electrostatic (HASEL) actuation, was introduced in 2018 by Eric Acome, Christoph Keplinger, and colleagues at the University of Colorado Boulder. A HASEL actuator places a liquid dielectric inside a sealed polymer pouch rather than a solid electrode-coated film; applying voltage pulls the pouch's flexible electrodes together, displacing the liquid and changing the pouch's shape, turning electricity directly into hydraulic pressure without a pump. Because the dielectric is a self-healing liquid, HASEL actuators survive local electrical breakdown that would permanently damage a conventional DEA[21]. The original paper demonstrated a two-unit planar HASEL actuator reaching 124 percent linear strain under a 700-gram load (about 114 kilopascals of stress), cycle efficiencies of roughly 20 to 30 percent, and a donut-style variant running past 1.08 million cycles at 18 kilovolts and 5 hertz with no noticeable loss of performance[22]. The Boulder spinout Artimus Robotics has commercialized HASEL modules, and in February 2026 announced a new generation claiming more than double the prior line's mechanical output, seeking robotics partners including in humanoid manipulation; HASEL had not been reported inside a shipping humanoid hand as of that announcement[23][24].

### Shape-memory alloy actuators

A shape-memory alloy (SMA) actuator, almost always nitinol, a roughly equal-parts nickel-titanium alloy, exploits a reversible solid-state phase change: below its transformation temperature the alloy sits in a soft martensite phase, and heating it, typically with resistive current through a wire, drives it into a stiffer austenite phase in which it snaps back to a pre-set shape, contracting[25]. Nitinol wire sustains high stress for its size, on the order of 200 megapascals, and an exceptionally high peak specific power, tens of kilowatts per kilogram, above most other families[26]. The tradeoffs are strain and speed: practical cyclic strain is usually limited to around 5 percent for useful fatigue life, versus roughly 20 to 40 percent for biological muscle, and thermal efficiency is poor, commonly cited under 10 percent, since most electrical input is lost as heat[25][26]. Bandwidth is capped by cooling speed, typically 1 to 5 Hz for simple wire without forced cooling; extreme miniaturization plus forced-air cooling has pushed a 2026 design using 25-micrometer nitinol wire bundles to 200 Hz, a record for electrically driven SMA wire actuators, though still a lab result rather than a deployed robot joint[27].

### Twisted and coiled polymer (supercoiled) actuators

The newest widely studied family began with an unusually accessible discovery. In 2014, Carter Haines, Ray Baughman, and colleagues at the University of Texas at Dallas showed that ordinary polymer fiber, including nylon fishing line and polyester sewing thread, could be twisted until it coiled into a tight spring, and that heating the coil (typically by resistive Joule heating) makes it untwist and contract along its length[28]. These are often called twisted and coiled polymer (TCP) actuators; a 2017 paper by Michael Yip and Gunter Niemeyer popularized the related term supercoiled polymer (SCP) actuator for the same family, with control methods for it[29]. Haines and colleagues reported contraction exceeding the roughly 20 percent in-vivo range of natural muscle, over one million cycles with low hysteresis, and showed that about 100 fishing-line muscles in parallel could lift roughly 0.8 metric tons[28]. Reported specific power is very high, commonly cited in the tens of kilowatts per kilogram, but thermal efficiency is extremely low, typically well under 2 percent, since Joule heating dissipates most input energy as waste heat[30]. As with shape-memory alloys, bandwidth is cooling-limited: heating is fast, but relaxation needs the fiber to shed heat, which in still air can take seconds per cycle without active cooling. TCP and SCP actuators remain mostly a research technology, popular in academic soft-robotics labs for low cost, but not yet reported in a shipping humanoid joint.

## Comparing artificial muscle technologies

Artificial muscles are usually compared on five rough axes: **strain** (contraction as a fraction of resting length), **stress** (sustained force per unit cross-sectional area), **specific power** (mechanical power per unit mass), **efficiency** (input energy converted to useful work rather than heat), and **bandwidth** (achievable cycle rate, often set by how fast the actuator resets, whether venting air, cooling a wire, or discharging a capacitor). The table compiles order-of-magnitude figures from a 2020 comparative review by Wei Liang and colleagues, cross-checked against the primary studies cited above[2]; performance varies by specific design, so these are illustrative ranges, not fixed specifications.

| Technology | Typical strain | Typical stress | Specific power | Efficiency | Bandwidth notes |
|---|---|---|---|---|---|
| Biological skeletal muscle (reference) | ~20-40% | ~0.1-0.35 MPa | ~0.05-0.28 kW/kg | ~20-40% | Benchmark other actuators are compared against[2][26] |
| Pneumatic McKibben / PAM | ~20-30% | ~0.1-1.2 MPa | ~0.5-10 kW/kg | ~25-49% | Up to ~150 Hz in small units; compressibility limits precision[2][5] |
| Hydraulic artificial muscle | ~20-30% | several MPa (5 MPa demonstrated) | comparable to or higher than pneumatic, same geometry | ~60% (vs ~25% pneumatic, same design) | Faster, stiffer than pneumatic; needs leak-free valves[10][11] |
| Dielectric elastomer actuator | up to >100% under prestrain | ~kPa to several MPa | ~3.6 kW/kg | up to ~80% | Up to ~1 kHz in some designs; needs kilovolt drive[2][19] |
| HASEL / electrohydraulic | up to ~100%+ (design-dependent) | ~0.1-0.3 MPa | ~0.6 kW/kg | ~20-30% | Demonstrated past 1 million cycles; kilovolt drive[2][22] |
| Shape-memory alloy (NiTi) | ~5% practical (fatigue-limited) | ~200 MPa | up to ~50 kW/kg (peak) | <10% | ~1-5 Hz without forced cooling; up to ~200 Hz miniaturized[2][26][27] |
| Twisted/coiled polymer (supercoiled) | up to ~50% | design-dependent | tens of kW/kg (order of magnitude) | typically <2% | Cooling-limited like SMA; fast heating, slow relaxation[2][28][30] |

Two patterns stand out. Several families beat biological muscle on specific power and strain by a wide margin; the deficit is almost always efficiency and, for the thermally driven families, bandwidth. No family wins on every axis, which is why developers keep experimenting with more than one approach rather than converging on a single winner the way the industry has converged on brushless motors for rotary joints.

## Tradeoffs versus conventional electromechanical actuation

The appeal of an artificial muscle is architectural as much as a matter of raw performance. A conventional electric joint bundles a motor, precision reducer, bearings, and encoder into one expensive, tightly toleranced assembly, with the reducer alone often close to half the joint's cost. A fluid-driven muscle can skip the reducer and precision-bearing stage entirely, since the contracting element pulls a joint directly, the way a tendon does. That is the bet behind Clone Robotics' and Sanctuary AI's approaches: trade a hard mechanical-tolerance problem (a near-zero-backlash gearbox) for a hard fluid-control problem (valves, pumps, and seals that are fast, small, and leak-free)[13][15].

That trade relocates complexity rather than eliminating it. Pneumatic systems need compressors and are typically noisy and imprecise. Hydraulic systems need a pump, reservoir, and valves that are fast and essentially leak-proof to be safe near people, historically unavailable at small scale or prohibitively expensive, the gap Sanctuary's and Clone's miniaturized valves target. Electrostatic families avoid pumps but introduce kilovolt-level voltage near people. Thermally driven families are electrically simple but bandwidth-capped by cooling, a poor fit for anything that must react as fast as human muscle. Boston Dynamics' history illustrates the industry's wariness of fluid power: its original Atlas ran on conventional hydraulic cylinders, not braided muscles, for roughly a decade before the company retired that platform in April 2024 for a fully electric successor, citing maintenance burden, leak risk, and plumbing complexity[31]. Every muscle technology now pitched for humanoids has to argue it solved the specific piece of that complexity, valve speed, voltage packaging, or thermal cycling, that made earlier fluid-power and thermal-actuator robots impractical.

## Use in humanoid robots

As of 2026, artificial muscles remain a minority approach in humanoid robotics next to motor-and-reducer joints. Clone Robotics and Sanctuary AI, detailed above, offer the most advanced demonstrated hardware, pneumatic-to-hydraulic Myofiber muscles across a full android body and hydraulic hand actuation respectively, with public demonstrations but no untethered, fully autonomous walking shown yet[13][16][17]. Festo's Fluidic Muscle is a decades-old industrial component rather than a humanoid-specific one, though its Bionic Learning Network has repeatedly used pneumatic muscles for compliant, human-safe grippers and arms[5][7]. Artimus Robotics sells HASEL modules commercially and is courting humanoid-manipulation customers, but had no reported production deployment as of 2026[23][24]. Shadow Robot Company is a useful counterpoint: its pneumatic air-muscle hand was a flagship product for roughly six years before the company judged electric motors better suited to precise finger control, a reminder these tradeoffs already pushed one prominent dexterous-hand developer away from pneumatic muscles before[9]. Shape-memory alloy and twisted-coiled polymer actuators, by contrast, remain confined mostly to smaller mechanisms and academic labs, held back by cooling-limited bandwidth in both cases.

One additional company deserves a caveat. Allonic, a Hungarian startup founded in 2025 that raised $7.2 million in pre-seed funding in early 2026, is developing a manufacturing process called 3D Tissue Braiding that weaves load-bearing fiber, actuators, tendons, and wiring directly around a robot's skeletal frame in one automated process, evoking musculoskeletal anatomy[32]. Public materials describe the braided structural approach in detail but do not specify whether its integrated actuators are pneumatic artificial muscles, and the company has not published actuator specifications, so its fit in this taxonomy could not be independently confirmed[33].

## Suppliers and landscape

| Organization | Technology | Notes |
|---|---|---|
| [Festo](/wiki/festo) | Pneumatic McKibben-type | Commercial Fluidic Muscle line; Bionic Learning Network demonstrators[5][7] |
| Bridgestone | Pneumatic and hydraulic McKibben-type | Original Rubbertuator; hydraulic high-power muscle with Tokyo Tech, 2017[3][12] |
| [Clone Robotics](/wiki/clone_robotics) | Pneumatic, transitioning to hydraulic ("Myofiber") | Full android bodies; moving to custom "Aquajet" hydraulic valves[15][16] |
| [Sanctuary AI](/wiki/sanctuary_ai) | Hydraulic (miniaturized valves) | Dexterous-hand actuation only, not full-body, as of 2026[13][14] |
| Artimus Robotics | Electrohydraulic (HASEL) | CU Boulder spinout; sells standalone modules to integrators[23][24] |
| Bayer MaterialScience (ex-Artificial Muscle, Inc., an SRI spinout) | Dielectric elastomer | Commercialized haptics (ViviTouch), not robot joints[18] |
| Shadow Robot Company | Pneumatic (historical) | Air-muscle hand from about 2004 to 2010, then switched to electric[9] |
| Allonic | Unspecified, integrated into a braided structure | Early-stage (founded 2025); actuator tech not independently verified[32][33] |

## See also

- [Actuator](/wiki/actuator)
- [Tendon-driven](/wiki/tendon_driven) actuation
- [Dexterous hand](/wiki/dexterous_hand)
- [Harmonic drive](/wiki/harmonic_drive)
- [Clone Robotics](/wiki/clone_robotics)
- [Sanctuary AI](/wiki/sanctuary_ai)
- [Humanoid robot](/wiki/humanoid_robot)

## References

1. Mirvakili, S. M. and Hunter, I. W. "Artificial Muscles: Mechanisms, Applications, and Challenges." Advanced Materials, vol. 30, 2018. https://advanced.onlinelibrary.wiley.com/doi/10.1002/adma.201704407
2. Liang, W. et al. "Comparative study of robotic artificial actuators and biological muscle." Advances in Mechanical Engineering, SAGE, 2020. https://journals.sagepub.com/doi/10.1177/1687814020933409
3. "Pneumatic Artificial Muscle." ScienceDirect Topics. https://www.sciencedirect.com/topics/engineering/pneumatic-artificial-muscle
4. "McKibben Pneumatic Artificial Muscle Robot Actuators: A Review." IEEE Xplore. https://ieeexplore.ieee.org/document/10089581/
5. "Fluidic Muscle DMSP." Festo product documentation. https://www.festo.com/media/catalog/202851_documentation.pdf
6. "Festo DMSP Series Fluidic Muscle." Motion World. https://www.motionworld.com/products/148939/festo-dmsp-series-fluidic-muscle
7. "BionicSoftArm" and "Bionic Motion Robot." Festo USA, Bionic Learning Network. https://www.festo.com/us/en/e/about-festo/research-and-development/bionic-learning-network/bionic-grippers-and-soft-robots/bionicsoftarm-id_68209/
8. "ExoHand." Festo USA, Bionic Learning Network. https://www.festo.com/us/en/e/about-festo/research-and-development/bionic-learning-network/bionic-grippers-and-soft-robots/exohand-id_33631/
9. "Dexterous Robotic Hands Part 1." Wevolver. https://www.wevolver.com/article/dexterous-robotic-hands-part-1-unraveling-the-history-technology-and-applications-of-dexterous-robotic-hands
10. Chambers, J. M. and Wereley, N. M. "Influence of hydraulic versus pneumatic working fluids on quasi-static force response of fluidic artificial muscles." Journal of Intelligent Material Systems and Structures, SAGE, 2021. https://journals.sagepub.com/doi/full/10.1177/1045389X21991927
11. "Development of a Hydraulic Drive High-Power Artificial Muscle." Tokyo Institute of Technology News, 2017. https://www.titech.ac.jp/english/news/2017/037286
12. "Development of a Hydraulic Drive High-Power Artificial Muscle." Bridgestone Corporation News, 2017. https://www.bridgestone.com/corporate/news/2017012601.html
13. "Sanctuary AI's Hydraulic Actuation Brings Human-level Dexterity To General-purpose Robots." Manufacturing AUTOMATION. https://www.automationmag.com/sanctuary-ais-hydraulic-actuation-brings-human-level-dexterity-to-general-purpose-robots/
14. "Back to Hydraulics? Clone and Sanctuary AI Boast Next-Gen Fluid-Powered Robots." Mike Kalil. https://mikekalil.com/blog/sanctuary-clone-hydraulics/
15. "Android." Clone Robotics. https://clonerobotics.com/android/
16. "Technology behind ghostly water-powered humanoid robot revealed." Interesting Engineering. https://interestingengineering.com/innovation/clone-robotics-technology-revealed
17. "Protoclone V1 is a sweating robot with 1,000 artificial muscles." R&D World. https://www.rdworldonline.com/protoclone-v1-1000-artificial-muscles-power-this-sweating-robots-human-like-moves/
18. "75 Years of Innovation: Artificial Muscle." SRI International. https://www.sri.com/75-years-of-innovation/75-years-of-innovation-artificial-muscle/
19. "A Review of Electrically Driven Soft Actuators for Soft Robotics." PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC9693343/
20. "Low-voltage and high-output dielectric elastomer actuators for untethered soft machines working at 200 volts." Science Robotics. https://www.science.org/doi/10.1126/scirobotics.ady9635
21. "CU Engineering Researchers Create Soft Robotic Muscles." University of Colorado Boulder College of Engineering, 2018. https://www.colorado.edu/engineering/2018/01/04/cu-engineering-researchers-create-soft-robotic-muscles
22. Acome, E. et al. "Hydraulically amplified self-healing electrostatic actuators with muscle-like performance." Science, vol. 359, 2018. https://imechanica.org/node/22096
23. "HASEL Actuator Technology." Artimus Robotics. https://www.artimusrobotics.com/technology
24. "Artimus Robotics launches next-generation artificial muscles." PR Newswire, February 2026. https://www.prnewswire.com/news-releases/artimus-robotics-is-launching-next-generation-artificial-muscles-and-seeking-partners-to-evaluate-the-technology-in-markets-requiring-dexterous-robotic-manipulation-302689585.html
25. "Ultimate Guide to Shape Memory Alloy Actuators in Robotics." GEE Nitinol. https://www.geenitinol.com/blog/ultimate-guide-shape-memory-alloy-actuators-robotics
26. "Shape Memory Alloy Actuators in Robotics." MDPI Actuators. https://www.mdpi.com/2076-0825/15/3/162
27. "Bidirectional rotational antagonistic shape memory alloy actuators for high-frequency artificial muscles." Scientific Reports. https://www.nature.com/articles/s41598-025-93209-9
28. Haines, C. S. et al. "Artificial Muscles from Fishing Line and Sewing Thread." Science, vol. 343, 2014, pp. 868-872. https://pubmed.ncbi.nlm.nih.gov/24558156/
29. Yip, M. C. and Niemeyer, G. "On the Control and Properties of Supercoiled Polymer Artificial Muscles." IEEE Transactions on Robotics, 2017. https://ieeexplore.ieee.org/document/7864464/
30. "Power-efficient low-temperature woven coiled fibre actuator for wearable applications." Scientific Reports. https://www.nature.com/articles/srep36358
31. "An Electric New Era for Atlas." Boston Dynamics official blog, April 2024. https://bostondynamics.com/blog/electric-new-era-for-atlas/
32. "Allonic raises $7.2M to build a foundation for advanced robotic hardware." Robotics 24/7. https://www.robotics247.com/article/allonic-raises-7.2m-to-build-a-foundation-for-advanced-robotic-hardware
33. "Allonic Secures $7.2M for 3D Tissue Braiding System." Mike Kalil. https://mikekalil.com/blog/allonic-3d-tissue-braiding-for-robots/