Step 1: a motor (DC, BLDC, McKibben pneumatic muscle or a servo) is mounted in the robot's body or forearm. Step 2: a drum/pulley is attached to the motor shaft, with a cable wound onto it. Step 3: the cable runs through guides (Bowden tubes or a pulley system) along the manipulator's structure to the distal joint. Step 4: at the joint, the cable terminates on one side of the rotation axis of a phalanx — pulling the cable generates torque. Step 5: because a cable can only pull, every joint needs either an antagonistic pair (two cables on opposite sides of the axis), or one cable with a return spring, or N+1 routing (N joints driven by N+1 cables with shared paths). Step 6: encoders on the motor shafts plus force sensors (Hall, strain gauges, BioTac) in the fingers close position and force control loops, with compensation calibrated for cable stretching and routing friction.
Direct drive (a motor in every joint) in a humanoid hand faces three design barriers: (1) a motor with meaningful torque does not fit inside a phalanx — the finger becomes thicker than a human one, (2) the distally-mounted motor mass increases moment of inertia, damping speed and raising energy cost, (3) motors inside a phalanx are exposed to impact and vibration, shortening lifespan. Tendon Drive solves all three — motors sit in the safe, large body/forearm, fingers are light and slender, and the cable transmission itself is passive and tolerant of momentary overloads (the cable slips or stretches instead of snapping a gearbox).
A motor physically remote from the moved joint — in the body, arm or forearm. Most often a BLDC with planetary gearbox (electric Shadow Hand variant), DC with reducer (research designs) or a McKibben pneumatic muscle (Shadow Air Muscle Hand, Festo BionicSoftHand).
Official
Passive transmission element: steel cable, UHMWPE polymer (Dyneema, Spectra), aramid (Kevlar) or carbon composite. Material choice balances breaking strength, axial stiffness, cyclic fatigue resistance and routing friction.
Official
A set of pulleys, Bowden tubes, guides and anchor points through which the cable travels from actuator to joint. Routing determines inter-joint coupling (coupled vs decoupled), efficiency (cable-routing friction) and packaging.
Official
Because a cable can only pull, every joint needs a mechanism to extend in both directions. The antagonistic variant (two cables on opposite sides of the axis, like biceps-triceps) offers active control of stiffness and force in both directions. The return-spring variant is simpler and lighter but gives up passive stiffness in one direction.
Motor-shaft encoders (shaft angle, not joint angle — a notorious pitfall), Hall-effect or potentiometers in joints (actual joint position), strain gauges/load cells on cables (tension), optionally tactile sensors (BioTac, Syntouch). The sensing layer compensates for cable stretch and friction.
Dyneema stretches by 1-3% under nominal tension; the motor encoder reports shaft position, not joint position. Without compensation, finger position error grows with force.
In long/curved Bowden paths, cable-guide friction creates hysteresis: the same joint angle requires different force depending on direction of motion. Impossible to compensate with open-loop control alone.
UHMWPE cables withstand static breakage well but are vulnerable to cyclic fatigue (especially on sharp pulley bends). Mean time between replacements in the Shadow Hand is ~10⁶-10⁷ cycles.
N+1 schemes and shared routing save actuators but make motion of one joint change the cable length serving another. Without a model this causes drift.
Rajko Tomović and Miodrag Rakić (University of Belgrade) demonstrate a 5-finger prosthesis with mechanical tendon control — the progenitor of modern prosthetic hands.
Kenneth Salisbury (Stanford) and JPL design a 3-finger tendon-driven hand with a 3-D kinematic and force model. The analytical foundations used to this day are established.
Foundation of the company that will turn tendon drive from research prototype into a commercial anthropomorphic-hand product — the Shadow Dexterous Hand becomes the industry standard for manipulation research.
OpenAI demonstrates reinforcement learning solving the Rubik's cube on a Shadow Dexterous Hand. For the first time, AI rivals humans in controlling a tendon-driven manipulator.
Tesla shows the Optimus humanoid hand with 11 degrees of freedom per hand and tendon drive. Tendon drive enters the consumer humanoid segment.
Vendors release 20+ DOF hands with dense tactile sensing and interfaces designed for VLA policy learning. Tendon drive becomes the de facto standard for humanoid hands trained from teleoperation.
Time complexity: Pasmo sterowania ograniczone do ~50-200 Hz (kabel UHMWPE) lub ~5-20 Hz (mięsień pneumatyczny). Space complexity: Liczba aktuatorów ≥ liczba stopni swobody (typowo 1.0-1.5×, dla schematu N+1 wynosi N+1).
The main bottleneck is not computation but identification of the cable-stretch model, friction hysteresis and inter-joint coupling for a given configuration. This requires per-unit calibration, and the model degrades as the cable wears.
Drives stiffness, breaking strength, cyclic life and friction.
Determines actuator count per DOF and whether the joint has active stiffness control.
Determines bandwidth, torque, compliance and power-supply scheme.
Drives inter-joint coupling and friction.
All tendons are active in parallel — classic dense control. The conditional aspect only appears when the policy decides which fingers to engage.
'Routing' in this concept is the literal physical guidance of the cable from motor to joint — unlike in MoE, where routing is a learned function. Here it is hard-wired into the hardware geometry.
Each finger is an independent control sub-loop, but inside a finger the joints are coupled through tendons.
Tendon Drive is itself a hardware concept; the control layer can run on any CPU/MCU. No special AI accelerators are required.
Position/force control loops run on regular CPUs or embedded MCUs (STM32, NXP) at 1-2 kHz.
Required only if the hand-control policy is a neural network (RL, VLA, imitation learning). The drive pattern itself is GPU-agnostic.
