Tendon-driven hands: the humanoid’s fingers moved to the forearm

Why the actuators left the hand

Your own fingers contain no muscle. The muscles that close your grip sit in the forearm and pull on tendons that run through the wrist and along the fingers. Humanoid builders keep converging on the same layout for the same physics. Fingers must be slim enough to reach into human-sized spaces and light enough to move fast; every gram of motor placed in a fingertip adds inertia that the wrist and arm then have to swing. Motors also produce heat, and a hand wrapped around an object has nowhere to dump it. Moving the actuators to the forearm frees the designer to size them for grip force rather than for whatever fits inside a phalange.

Tesla’s Optimus V3 patents, published in 2026, are the clearest statement of this shift. Earlier Optimus generations carried actuators in the hand itself — the design Elon Musk described as producing “giant hands that look weird”. The patents relocate them into the forearm and describe three thin tendons per finger running from forearm actuators through the wrist into channels inside the finger bones, for roughly 22 degrees of freedom across hand and wrist. Rob Knight, the hand builder behind the Robot Nano Hand project, put the engineering logic bluntly in our interview: at finger scale, tendons still beat tiny gearboxes.

Three ways to drive a finger

Tendons are not the only answer, and the market currently runs three architectures side by side. Full remote tendon drive is the high-dexterity path: Optimus V3, 1X’s NEO hand and the Shadow Dexterous Hand all pull 20 or more joints from forearm-mounted actuators. Linkage drive is the workhorse path: Inspire Robots’ RH56 series, the merchant hand fitted to many Chinese humanoids including Unitree G1 and H2 variants, packs six motors into the hand itself and drives twelve joints through rigid linkages — fewer degrees of freedom, but cheap, sealed and robust. And hydraulics is the outlier: Sanctuary AI’s Phoenix uses miniature hydraulic valves to actuate hands with roughly 20 degrees of freedom, trading system complexity for very high force density at the fingertip.

Figure sits deliberately outside this taxonomy in public: the company has not disclosed the Figure 03 hand’s drive mechanism or DoF count, and talks instead about what the hand senses — fingertip tactile sensors that detect forces down to about three grams, a wide-angle palm camera for close-range vision, and softer adaptive fingertips. That is its own bet: that sensing and compliance, not raw joint count, decide real-world manipulation.

The tendon is now a component, and it wears

Once the actuators leave the hand, the tendon becomes the load path — and the failure path. The materials shortlist is short. Ultra-high-molecular-weight polyethylene fibre (UHMWPE, sold as Dyneema) dominates because it is stronger than steel by weight, flexible around small pulleys and low-friction, but it creeps: studies of Dyneema under cyclic loading show progressive strain accumulation, which in a hand means tendons that slowly lengthen and lose calibration. Aramids such as Kevlar and Technora creep far less but abrade faster at pulley contacts; liquid-crystal polymer and ePTFE cover niche friction and routing cases; steel cable survives heat but fatigues at the tight bend radii a finger demands. Specialist cord makers now market robotic-tendon lines that are engineered against exactly these trade-offs — routing conditions, bend radius, pulley geometry and expected motion cycles.

Routing is the other half of the problem. Every tendon crossing the wrist changes length slightly as the wrist moves, and neighbouring tendons rub, so an uncompensated design has fingers that twitch when the wrist rotates. Tesla’s patents attack this directly with what amounts to a cable router in the wrist: tendon bundles leave the forearm in a lateral stack and are re-ordered into a vertical stack through a transition zone shaped to cut friction, minimise crosstalk between channels and preserve tension during combined wrist movements. The patents’ stated design goals — planar geometries, few exotic parts, tolerance of millions of duty cycles — read as an admission that in a factory robot, tendons are a wear item that must be cheap to make and cheap to replace.

Degrees of freedom: where each hand lands

Degrees of freedom are not dexterity — control software, sensing and tendon quality decide what a hand can actually do — but they set the ceiling, and the published numbers show how close the leading tendon hands now sit to the human hand’s commonly cited 27. They also show the gap down to the linkage-driven hands that most deployed humanoids actually wear today: an Inspire-style hand actuates 6 of its 12 joints, so many finger joints are coupled and cannot be commanded independently.

What tendon hands mean for the supplier landscape

The hand is the one humanoid subsystem that already behaves like a merchant market. The OEMs building tendon hands keep them in-house — Tesla, 1X and Figure all treat the hand as core IP — but everyone else buys. That is why the interesting supplier stories cluster in three places.

• Merchant hand makers. Inspire’s linkage-driven RH56 series is the de facto standard fitment on Chinese humanoids, sold in variants matched to Unitree’s G1 and H2. Above it sits a fast-growing tier of dexterous-hand specialists — Shadow at the research end, and newer entrants such as Mimic, Wuji and Rochu that we track in the Hand Tracker.
• Tendon and cordage suppliers. Braided UHMWPE, aramid and LCP cords engineered for robotic tendons are becoming a catalogue item, with cord makers publishing robotics-specific selection guides. When a hand has 20+ tendons and each is a wear part, tendon cord is a recurring-revenue component, closer to brake pads than to gearboxes.
• The service economy. A tendon that creeps needs re-tensioning or replacement; the Optimus patents’ emphasis on million-cycle durability and simple manufacturable geometry is aimed squarely at keeping that maintenance cheap. Fleet operators should ask every hand vendor two questions: what is the tendon’s rated cycle life, and how long does a tendon swap take.

FAQ

Sources