Hands and touch

A pair of Optimus-class hands costs about US$9,500, around 17% of the robot's parts bill, more than any leg segment estimate. They hold 16 to 20 tiny motors each, they touch the world thousands of times a day, and the software to use their sense of touch is still early. Most robots that earn money today use a two-finger clamp or a suction cup instead. This chapter shows you why.

The gripper spectrum

Anything a robot uses to hold things is an end effector, and most of them are a gripper of some kind. They run from simple to complex, and the simple ones do most of the paid work.

• Parallel grippers: two flat jaws that slide together. One motor, cheap, strong, easy to control and easy for a learning model to use. Dyna's napkin-folding stations and most research arms use them; Agility's Digit uses swappable grippers on standard mounting flanges verified.
• Suction cups: a vacuum pad that sticks to flat surfaces. Boston Dynamics' Stretch unloads about 500 parcels an hour with one verified, and Galbot's G1 has shipped with a suction cup on one arm and a parallel gripper on the other contested.
• Soft grippers: rubbery fingers that bend around fruit, bread or bags without crushing them.
• Underactuated grippers: jointed fingers with fewer motors than joints. The fingers wrap until they touch, then stop, so they fit odd shapes without a model of the object.
• Dexterous hands: five fingers, 6 to 22 motor-driven joints, often with touch sensors in the fingertips estimate. They can use tools built for people, and they cost and break like it.

Two words for how a hand holds things will come up in every demo. A power grasp wraps the palm and fingers around an object, like holding a hammer: strong, forgiving, low precision. A precision grasp holds with the fingertips, like picking up a coin: weak, exact and dependent on touch. A robot that only does power grasps can move totes and parts. A robot that does precision grasps can plug in a connector or fold a shirt, and that's the skill buyers pay a premium for. Both are forms of grasp, and everything a hand does after the grasp, turning, sliding, pressing, is manipulation.

Counting joints: DoF and actuated DoF

A human hand has about 27 degrees of freedom estimate, meaning 27 independent ways it can move. Robot hand makers quote two numbers, and both matter. DoF counts the joints that can bend. Actuated DoF counts the joints that have their own motor. A finger whose middle and tip joints are linked to bend together has three DoF and one actuated DoF.

A fully actuated hand can do more in-hand tricks, and it needs more motors, more wires, more heat and more control. An underactuated hand is cheaper, lighter and more robust, and it wraps naturally, which is often what a task needs. Many hands mix the two: Unitree's Dex5-1 drives 16 of its 20 joints with motors and lets 4 follow passively.

In the finger, the palm or the forearm

Where the motors sit decides how the hand looks, what it weighs at the end of the arm and how you repair it. There are three layouts.

• Motors in the fingers. Tiny coreless motors, 6 to 8 mm across, sit inside each finger segment with a micro gearbox or screw estimate. The hand is self-contained and each finger can be a replaceable module. Unitree's Dex5-1 uses coreless motors and makes each finger replaceable verified. The cost is fat fingers and heat right where the hand touches things.
• Motors in the palm. Small linear actuators in the palm push the fingers through rods. Inspire Robots' RH56 series drives its fingers with micro servo "electric cylinders" verified, and Unitree offers Inspire hands on G1 EDU versions. This is the linkage-driven hand: stiff, precise and relatively cheap, with fewer independent joints.
• Motors in the forearm. Motors sit in the forearm and pull tendons, cables that run through the wrist to the fingers, like yours. This tendon-driven hand has slim, light fingers. 1X NEO puts its hand motors in the forearm verified, and Optimus Gen 3 reportedly uses about 25 forearm actuators per arm pulling tendons for a 22 DoF hand unverified.

The motor that makes all of this possible is the coreless motor, a motor whose spinning part is a thin, self-supporting copper coil with no iron inside, so it starts and stops almost instantly. A dexterous hand uses 16 to 20 of them, so a two-handed humanoid carries 30 to 40 in its hands alone estimate. Imported units cost about CNY 2,000 each against about CNY 1,000 for Chinese ones, with a volume target near CNY 600 estimate. Maxon and Faulhaber are the Swiss and German references; Moons' in Shanghai is the leading Chinese maker, and Zhaowei in Shenzhen makes micro gearboxes and drives for hands unverified.

Tactile sensing: giving fingertips a sense of touch

Robots can see, and they can feel force at their joints. Most can't feel where on a finger they're touching something, or whether it's starting to slip. A tactile sensor fixes that: a grid of tiny pressure cells, called taxels, on a fingertip or palm that reports where contact happens and how hard, sometimes including sideways shear. Spread over a larger surface it's called electronic skin.

Costs are falling fast in China. Tactile arrays run about US$0.4 to 1.2 per square millimetre estimate. PaXini, a Shenzhen startup founded in 2021, sells magnetic fingertip sensors from about US$49 estimate and claims CNY 199 or lower against the historic price of about CNY 100,000 for an imported fingertip unverified. It raised more than CNY 1 billion in a B round in March 2026 at a valuation above CNY 10 billion verified, and cumulative funding passed CNY 4 billion in September 2026 estimate. It claims 80% penetration among leading humanoid and hand makers unverified.

The sensor hardware is the easy half. The hard half is software that uses the signal. Figure's Helix 02 model feeds palm cameras and fingertip tactile readings into its policy verified, and Sharpa's CraftNet claims to be a vision-touch-language-action model company claim.

Wrist force and torque sensors

Between the arm and the hand, most serious manipulators carry a six-axis force/torque sensor. It's a machined metal ring with strain gauges that bend microscopically under load, and it reports three pushes (forward, sideways, up) and three twists around those axes. It tells the robot the total load on the hand: how heavy the box is, whether the peg is jammed, whether the arm just hit something. The same sensor sits in some ankles.

China's humanoid makers bought about 12,300 six-axis sensors in 2025, up 510%, with domestic brands holding 92.8% of that market and Landian Touch alone 72.6% estimate. Prices: CNY 15,000 to 38,000 for a wrist unit, CNY 5,000 to 10,000 for an ankle, more than CNY 50,000 for hand-grade units, with a stated industry target of about CNY 1,000 estimate. Western references are ATI (US$3,000 to 15,000) and newer entrants like Bota (US$1,000 to 3,000) estimate.

Why hands are the hardest part

1. Packaging. A hand has to fit 16 to 20 motors, gearboxes, sensors, wiring and cooling into the volume of a human hand, and hold a few kilograms. Dex5-1 weighs 1.1 kg and lifts 3.5 kg palm down verified. Every gram sits at the end of a long arm and multiplies the load on the shoulder.
2. Durability. Fingertips hit things all day. Tiny gears, thin cables and soft skins wear fastest. CNN's reporting on China's robot rental market found that dexterous hands suffer from high cost, poor durability and short life verified.
3. Cost. The hands are the single most expensive body region of an Optimus-class robot, more than any leg segment estimate.
4. Contact simulation. Robots learn a lot in simulation, and simulators handle walking far better than touching. The gap is mostly friction, compliance and contact verified. World models make video look real; contact and force still don't transfer well.
5. Data. In-hand handling of cables, cloth and soft objects remains unsolved at human level estimate. Labs are turning to human video instead of robot data: Dyna-2 pretrained on more than a million hours of head-camera video with 3D hand-pose tracking verified, and Generalist's GEN-1 on 500,000+ hours of physical-interaction data from wearables verified. You can't record touch from a video, which is one reason tactile hands lag.

Hands on named robots

Notice who skips the dexterous hand. Agility's Digit uses clamp-style grippers because totes need a clamp, and Agility spends its money on uptime analysis. Apptronik's Apollo has mostly used grippers and has published no dexterous hand spec unverified. Boston Dynamics' production Atlas shows three-finger grippers in demos unverified. The robots with the most paid hours in warehouses are the ones with the simplest hands.

Hands break first

The evidence that hands and wrists fail first is thin but consistent. Figure named the forearm as the top hardware failure point of Figure 02 across 11 months at BMW: tight packaging, a 3 DoF wrist and thermal limits, with a distribution board and moving cables feeding the wrist motors. Figure 03 removed the board and the moving cables, and each wrist motor controller now talks straight to the main computer verified. On a Unitree G1, a lab reported a broken wrist pitch gear and an overheating wrist motor after violent learned motions owner report, and another owner found that adding a heavier Inspire hand made the shoulder motor overheat within minutes owner report. A hand's weight becomes the arm's heat problem.

Makers already design for this. Unitree sells Dex5 with each finger replaceable verified. NEO's hands are rated waterproof while the body isn't unverified. Treat a robot hand like a tyre: a wear part with a price, a life in cycles and a stock level.

Terms in this chapter