Tendon-driven actuation

Tendon-driven actuation (also called cable-driven actuation) is a method of moving a robot's joints by pulling on thin, flexible cables (the "tendons") that run from motors mounted somewhere else on the machine, over pulleys or through flexible sheaths, to the joint being moved. It is the mechanical opposite of putting a motor directly at each joint: the "muscles" (motors) sit remotely, usually in the palm or forearm, and only lightweight cables reach the fingers. This mirrors the human hand, where most of the muscles that curl the fingers actually live in the forearm and pull on long tendons that cross the wrist. Because it keeps the moving parts light and slim while packing many actuators into a compact space, tendon drive has become the dominant approach for high-degree-of-freedom dexterous hands and humanoid robot hands, and it is widely used elsewhere in robotics, from surgical tools to exosuits ^[1][2].

How tendon-driven actuation works

The core idea is force transmission at a distance. A motor rotates a spool or drives a small linear actuator, which takes up or releases cable. The cable is routed to a distant joint, where it wraps around a pulley or is anchored to a link. Pulling the cable applies a torque that rotates the joint, and releasing it lets the joint move back. Guiding the cable requires idler pulleys, capstans, or a sheath so it follows the intended path without snagging or slipping off ^[1][2].

Two routing styles dominate. In open routing, the cable runs over exposed pulleys and is well suited to short, mostly straight paths inside a finger or a rigid structure. In a Bowden cable, the inner wire slides inside an outer sheath, so force can be sent along a curved or moving path, for example from a fixed motor pack across a wrist that is itself rotating. Bowden routing buys freedom in where the motors live, but the price is friction: every bend presses the wire against the sheath, and tight curves multiply the losses according to the capstan effect, so designers keep bend radii as large as they can ^[4].

The biological analogy is exact enough to be useful. In the human hand, the large flexor muscles sit in the forearm and pull on tendons that run through sheaths and pulleys at the wrist and along each finger. Engineers copy this "extrinsic" scheme directly: motors in the forearm, tendons to the fingers ^[1][5].

Why hands use it

For an anthropomorphic hand, tendon drive solves a packaging problem that has no clean alternative. A human-scale finger is only about a centimeter thick, yet a capable hand needs on the order of 15 to 25 controllable joints. Fitting a motor and gearbox strong enough for a useful pinch into each finger segment is difficult, and doing it 20 times over is close to impossible at human size ^[1][6].

Tendon drive sidesteps this in several ways at once:

• Low distal inertia. With the motors moved back to the palm or forearm, the fingers themselves are little more than light links and pulleys. Low mass at the fingertips means faster, safer motion and less momentum to control. One industry analysis of optimized tendon hands puts the reduction in hand mass at roughly 30 to 40 percent and the drop in moment of inertia near 45 percent versus fully integrated actuation ^[5].
• Packing many degrees of freedom into a small volume. Cables are thin, so a slim finger can carry several of them, letting a hand reach 20 or more independently controllable joints at human scale ^[1][5].
• Room for bigger motors. The forearm has far more space than a fingertip, so designers can use larger, more powerful actuators than would ever fit inside a finger, then gear them down for high fingertip force ^[5].
• Speed and compliance. Short, light fingers respond quickly, and a cable transmission is naturally backdrivable and springy, which gives the inherent compliance that makes contact-rich grasping safer ^[3][5].

These properties are why nearly every high-DOF research and product hand that aims to look and behave like a human hand has used tendons ^[1].

Mechanics: how tendons move joints

A cable can only pull, never push, so every tendon needs a way to bring its joint back. How designers solve that defines the main architectures ^[1][3].

Antagonistic pairs (2N tendons)

The most human-like arrangement puts two tendons on each joint, an agonist and an antagonist, that pull in opposite directions. Controlling N joints this way takes 2N tendons and, in the fully independent case, 2N motors. The reward is active control in both directions plus the ability to co-contract both tendons at once to stiffen or soften the joint, exactly as a person tenses opposing muscles. The cost is a large number of actuators and cables to route and keep in tension ^[3].

N+1 tendons

A classic result in hand design (associated with the Salisbury and Utah/MIT work) is that N joints in a chain can be controlled with N+1 tendons rather than 2N, using one shared return tendon. This cuts the actuator count sharply, but it couples the joints together, so the motion of one affects the tension available to the others, which complicates control ^[1][3].

Unidirectional with a return spring

The cheapest scheme uses a single active tendon per joint to flex it and a spring (or an elastic element) to extend it when the cable relaxes. This halves the tendon and motor count, but the spring, not a motor, now sets the extension force, so the hand pushes back weakly in that direction. Many low-cost and underactuated hands accept this trade ^[1].

Beyond the tendon count, three details make or break a design. Pretension keeps cables taut so motion transfers immediately: too little and the hand feels slack and imprecise, too much and it accelerates wear and can buckle the structure. Pulley and idler geometry sets the moment arm at each joint and how cleanly the cable turns corners. And coupling and underactuation are used deliberately: routing one motor's tendon across several joints lets a hand conform around an object with fewer actuators, the principle behind most adaptive grippers ^[3][4].

Tendon materials

The tendon itself is a surprisingly demanding component, because it must be strong, thin, low-friction, and resistant to millions of bend cycles.

The workhorse material is ultra-high-molecular-weight polyethylene (UHMWPE), sold as Dyneema or Spectra. It offers a specific strength reported at more than 10 to 15 times that of steel by weight, at roughly one-seventh the density, with good abrasion resistance and low friction ^[6][7]. Material data cited in dexterous-hand industry analyses claims UHMWPE keeps more than 95 percent of its strength after 2 million tensile cycles, whereas ordinary polyester fiber loses about 30 percent of its performance after 1 million cycles ^[5][6].

UHMWPE's weakness is creep: under sustained load it slowly stretches, so a tendon set to the right length can go slack over time and need recalibration. Testing of Dyneema rope under cyclic load has shown elongation of more than 9 percent that continued to grow even past 500 cycles, though pre-tensioned, heat-treated Dyneema can be brought under 1 percent ^[8][9]. Fiber tendons also add bending friction when they wrap ball-bearing pulleys. For those reasons some builders switch to steel or tungsten cable, which is dimensionally stable and predictable but heavier, stiffer to route, and prone to fatigue cracking at tight bends ^[8]. The choice is a direct trade between the low weight and friction of polymer fibers and the stability of metal wire.

Trade-offs and the durability problem

Tendon drive buys its packaging advantage at a real cost in precision and upkeep ^[1][2]:

• Stretch and creep introduce a variable offset between what the motor commands and where the joint actually sits.
• Friction and hysteresis, worst in Bowden routing, mean the force at the fingertip lags and differs from the force at the motor, so the hand does not return to exactly the same place for the same command. This directly degrades position and force accuracy ^[4].
• Routing complexity grows fast with the tendon count; each cable needs a clean path, correct pretension, and its own failure mode.
• Maintenance is heavy. Tendon systems are sensitive to tension and layout, need periodic recalibration, and when a tendon breaks the hand often must be partly disassembled to rethread it ^[6].

Durability is the headline limitation, and it is a well-documented failure mode. Industry analysis of hand transmissions notes that steel-wire tendons have a working life of only "thousands to tens of thousands of cycles" before they must be replaced, and that polymer tendons age and need frequent recalibration as they wear and slacken ^[6]. In other words, a traditional tendon-driven hand can begin to fail on the order of roughly 10,000 grasping cycles, which for a hand meant to work an industrial shift is far too few. Vendors now treat cycle life as a headline spec. As a reference point for what "good" looks like, the startup Xynova states that its dexterous hand passed a 10,000 open-close test with a failure rate below 0.1 percent, a figure best read as an industry or vendor reliability target rather than an independently audited result ^[10].

Xynova makes the strongest published durability claims for a tendon hand. The company says the tendon transmission in its Flex 1 exceeded 1,000,000 cycles under rated load, and reports component-level aging tests of about 1,800,000 cycles for the tendon-rope assembly and more than 5,000,000 cycles for its in-house micro electric cylinder, with key components measured beyond 1,500,000 cycles under customer durability testing ^[10][11]. For the hybrid Flex 2, Xynova cites a service life on the order of 1,000,000 open-close cycles ^[13][14]. These are vendor claims, not independent test results, and should be read as such, but if borne out they represent roughly a hundredfold improvement over the traditional tens-of-thousands-of-cycles figure.

How tendon drive compares with other actuation

Tendon drive is one of several ways to power a hand, and the right choice depends on how many independent joints are needed, how much the fingers must weigh, and how much control precision and maintenance the application can tolerate ^[5][6].

Scroll sideways for more →

Direct drive embeds a micro-actuator in each finger segment. It removes cables entirely, so there is no slack, creep, or recalibration, and it can run at high control rates near 1 kHz. The catch is that every joint's strength and count are limited by what a motor that small can deliver, and the fingers carry all that motor mass themselves ^[21][22]. Linkage and gear coupling is mechanically robust and cheap: a handful of motors drive many coupled joints through rigid links, which is ideal for reliable industrial grasping but gives up independent control of each joint ^[23]. Hydraulic actuation, used by Sanctuary AI in its Phoenix robot's 21-DOF hands, trades cables for coin-sized valves and pressurized fluid, reaching high power density and fine force control; Sanctuary says its miniaturized valve components have been tested past two billion cycles without leakage or wear ^[24][25]. Artificial muscle, the route taken by Clone Robotics with its water-driven Myofiber actuators in the Clone Alpha android, is the most biomimetic of all: the company reports its fibers can contract about 30 percent in under 50 milliseconds and have survived 650,000 test cycles, though the technology is younger and less proven than tendon drive ^[26]. In practice many "tendon" hands are actually hybrids; the Xynova Flex 2, for example, combines tendon rope with direct motor drive ^[13].

Notable tendon-driven hands and systems

Scroll sideways for more →

The Shadow Robot Dexterous Hand is the reference example. First sold in 2005, it uses 24 degrees of freedom (20 actuated) and drives its fingers with tendons from motors or pneumatic air muscles mounted in the forearm, wrapped in more than a hundred sensors sampled at 1 kHz. It became the platform for landmark manipulation research, including OpenAI's Dactyl in-hand cube reorientation and a durable three-fingered variant, the DEX-EE, built with Google DeepMind for reinforcement-learning research ^[15][16].

The most consequential recent adopter is Tesla. Patents for the Optimus Gen 3 hand describe a redesign that roughly doubled hand dexterity to about 22 degrees of freedom by relocating around 25 actuators per side into the forearm and driving each finger with three thin control cables routed through the wrist into the finger phalanges, so that pulling a cable bends the corresponding joint. Tesla frames this as directly copying human anatomy: the forearm is the muscle, the hand is mostly tendon and bone ^[17][18]. The Figure 03 humanoid, revealed in October 2025, is another compact high-DOF hand with three-gram fingertip tactile sensing and a palm camera, although Figure has not publicly detailed whether its fingers are tendon-driven ^[19][20].

On the commercial frontier, Xynova, a 2024 Chinese startup backed by CATL and Xiaomi, describes its Flex 1 as the world's first fully self-developed, mass-producible, high-DOF tendon-driven hand: a 25-DOF, 380-gram hand that it says can handle loads above 30 kg and deliver over 20 N of fingertip force, completing an open-close cycle in about 0.6 seconds ^[11][12]. Not every strong hand uses tendons, and the contrast is instructive. The Wuji Hand deliberately rejects tendons in favor of direct-drive micro-actuators embedded in each finger, arguing that removing cables removes the calibration drift and slack that plague tendon systems ^[21][22]. The Inspire Robotics RH56 instead uses integrated linkage drive, coupling six motors to twelve joints for a rugged, low-maintenance industrial hand ^[23].

Beyond hands

The same principle scales well past hands wherever motors need to be kept away from the working end ^[2].

• Continuum and surgical robots. Long, snake-like tools bend by pulling cables that run along a flexible backbone; coordinating several cables steers the tip in any direction. Cable drive is the most common way to actuate continuum robots, and Bowden-cable tendons let surgical instruments articulate a tip inside the body while the motors stay outside the patient ^[27][28].
• Exosuits and exoskeletons. Soft wearable robots route Bowden cables from a waist-mounted motor pack to the ankles, hips, or elbows, delivering assistive force without heavy motors on the limb, which is exactly the low-distal-inertia argument applied to a person ^[29].
• Legged and full-body robots. Tendon and cable transmissions are used to place leg and arm actuators near the body's center of mass, cutting limb inertia for faster, more dynamic motion, the same reasoning that motivates tendon hands.

ELI5

Imagine a puppet. Instead of putting a tiny motor inside each of the puppet's little fingers, you put big motors up in the puppet's arm and run strings from those motors down to the fingers. Pull a string and a finger curls; let it go (or let a little spring pull it back) and the finger straightens. That way the fingers stay light and thin like real fingers, and the strong "muscles" hide up in the arm, just like in your own hand, where the muscles that move your fingers are actually in your forearm. The tricky part is that strings can stretch, rub, and wear out, so the hard engineering is making strings that last for millions of pulls.

See also

• Dexterous hand
• Humanoid robot hands
• Tactile sensing
• Degrees of freedom
• Harmonic drive
• Shadow Robot
• Xynova Flex 1
• Wuji Hand
• Tesla Optimus Gen 3
• Sanctuary AI Phoenix
• Clone Robotics Alpha

References

1. "A comprehensive review of dexterous robotic hands: design, implementation, and evaluation," Bioinspiration and Biomimetics (IOP Publishing), 2025. https://iopscience.iop.org/article/10.1088/1748-3190/ade7e1 ↩
2. "Tendon Driven Robotic Hands: A Review," International Journal of Mechanical Engineering and Robotics Research, 2015. https://www.ijmerr.com/uploadfile/2015/0409/20150409024829230.pdf ↩
3. Abdallah, Platt and Wampler, "Decoupled Torque Control of Tendon-Driven Fingers with Tension Management," International Journal of Robotics Research, 2013. https://academicweb.nd.edu/~cwample1/Preprints/AbdallahPlattWampler_IJRR_32-2_2013.pdf ↩
4. "Antagonistic Bowden-Cable Actuation of a Lightweight Robotic Hand: Toward Dexterous Manipulation for Payload-Constrained Humanoids," arXiv:2512.24657, 2025. https://arxiv.org/html/2512.24657v1 ↩
5. "Tendon-driven Solution for Dexterous Hands: Key Path and Ideal Choice for High Dexterity and Humanoid Manipulation," 36Kr, 2025. https://eu.36kr.com/en/p/3562387302701191 ↩
6. "腱绳：灵巧手打通'灵巧'的下一个破局点" (Tendon rope: the next breakthrough point for dexterous hands), Gelonghui, 2025. https://m.gelonghui.com/p/2536942 ↩
7. "UHMWPE vs Dyneema: Are They Really the Same Rope?" Orion Cordage. https://feeds.orioncordage.com/blog/uhmwpe-vs-dyneema ↩
8. Friedl et al. (DLR), "FRCEF: The New Friction Reduced and Coupling Enhanced Finger" (Dyneema creep versus steel tendons). https://elib.dlr.de/100496/1/FRCEF.pdf ↩
9. "Investigating the Strain Behaviour of Dyneema under Cyclic Loads," 2020. https://www.researchgate.net/publication/347169803_Investigating_the_Strain_Behaviour_of_Dyneema_under_Cyclic_Loads ↩
10. "万台订单，量产开启：曦诺未来Xynova如何定义灵巧手的'交付时代'" (Xynova mass production and durability testing), Zhihu, 2026. https://zhuanlan.zhihu.com/p/1993040874209297567 ↩
11. "CATL, Xiaomi Back Robotic Hand Startup Xynova Future in Angel Round," ChinaBizInsider, 2025. https://chinabizinsider.com/catl-xiaomi-back-robotic-hand-startup-xynova-future-in-angel-round/ ↩
12. "Chinese Robotics Startup Xynova Raises Over US$13 Million in Angel Round for High-DOF Dexterous Hands," TMTPost, 2025. https://en.tmtpost.com/post/7822011 ↩
13. "Chinese firm unveils Flex-2 arm to advance humanoid robots toward more versatile tasks," Interesting Engineering, 2026. https://interestingengineering.com/ai-robotics/chinas-new-robotic-hand-combines-hybrid-actuation-for-smarter-robot-manipulation ↩
14. "'Xynova' Defines New Height of Dexterous Operation with Flex2 after Securing Hundreds of Millions of Yuan in Series A Financing," 36Kr, 2026. https://eu.36kr.com/en/p/3829065404211456 ↩
15. "Shadow Dexterous Hand Series," Shadow Robot Company. https://shadowrobot.com/dexterous-hand-series/ ↩
16. "Shadow Robot DEX-EE hand takes manipulation to next level" (Shadow Robot and Google DeepMind), The Robot Report, 2024. https://www.therobotreport.com/shadow-robot-dex-ee-hand-takes-manipulation-to-next-level/ ↩
17. "Tesla Optimus V3 hand and arm details revealed in new patents," Teslarati, 2026. https://www.teslarati.com/tesla-optimus-v3-hand-arm-details-revealed-new-patents/ ↩
18. "Tesla Optimus V3 Robot Hand Patent Reveals Tendon-Driven, Human-Like Finger Architecture," Gear Musk, 2026. https://gearmusk.com/2026/04/17/optimus-v3-hand-patent/ ↩
19. "Introducing Figure 03," Figure AI, 2025. https://www.figure.ai/news/introducing-figure-03 ↩
20. "Figure 03 Is The Robot in Your Kitchen," TIME, 2025. https://time.com/7324233/figure-03-robot-humanoid-reveal/ ↩
21. "Experts Weigh In on Wuji Tech's Robotic Hand: A 'Remarkably Robust' Direct-Drive Contender," Humanoids Daily, 2025. https://www.humanoidsdaily.com/news/experts-weigh-in-on-wuji-techs-robotic-hand-a-remarkably-robust-direct-drive-contender ↩
22. "Wuji Tech enters robotic hand race with bold direct drive design," Tinkimo, 2025. https://tinkimo.com/wuji-tech-enters-robotic-hand-race-with-bold-direct-drive-design/ ↩
23. "Integrated linkage-driven dexterous anthropomorphic robotic hand," Nature Communications, 2021. https://www.nature.com/articles/s41467-021-27261-0 ↩
24. "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/ ↩
25. "Sanctuary AI's Phoenix Now Has a Robotic Hand with 21 DoF," Analytics India Magazine. https://analyticsindiamag.com/ai-news-updates/sanctuary-ais-phoenix-now-has-a-robotic-hand-with-21-dof/ ↩
26. "Clone demos humanoid robot hand with human-like grip strength" (Myofiber artificial muscle), Interesting Engineering. https://interestingengineering.com/ai-robotics/clone-demos-creepy-humanoid-hand ↩
27. "Unlocking the Potential of Cable-Driven Continuum Robots: A Comprehensive Review and Future Directions," Actuators (MDPI), 2024. https://www.mdpi.com/2076-0825/13/2/52 ↩
28. "Tendon-Driven Continuum Robot for Endoscopic Surgery: Preclinical Development and Validation of a Tension Propagation Model," 2015. https://pmc.ncbi.nlm.nih.gov/articles/PMC4569018/ ↩
29. "Sensorless model-based tension control for a cable-driven exosuit," arXiv:2406.18412, 2024. https://arxiv.org/html/2406.18412 ↩