Hands

Before that can be tested we have to say what a hand is, because the “22” on a spec sheet can mean three different things.

Start with your own. The human hand has 27 bones: 8 in the wrist, 5 in the palm and 14 in the fingers paperhere an encyclopedia. The muscles that curl the fingers are mostly not in the hand. They sit in the forearm and pull on long tendons paperthe same. Make a fist and watch your forearm move.

A robot hand has three numbers, and they are rarely equal. Joints are the hinges. Driven joints are the hinges that can be moved independently of the others. Motors are what moves them. LEAP v2 has 16 joints moved by 8 motors paper. Inspire’s RH56 drives 12 joints with 6 actuators maker’s claim. One hand in Tesla’s patent application has 22 finger degrees of freedom driven by 17 actuators maker’s claim. When one motor moves several joints, something else decides how they share the motion. That is chapter 3.

View B · fingers, close up Render of a simulation

Three things stand out. Most cells are empty. Of the eighteen hands that others built, six publish a fingertip force, and no two say they use the same test, so we would not rank them. Mass depends on where the motors are: Shadow’s 4.3 kg includes a forearm full of motors, and Wuji’s 580 g is the whole hand with its motors inside. And six motors are already the bottom of the market. Inspire’s hand is exactly that, and Inspire took 96.19 % of what Unitree spent on bought-in hands in the first nine months of 2025 pressfrom the filing. Six motors are enough to sell.

Are they enough to work? We built a six-servo hand in simulation, with 16 joints and 5 tendons. It closed into a fist and kept hold of a 150 g bar after we let go of the bar and shook it simulationchapter 12. Then we asked the model where its fingers could reach. With six driven joints the thumb touched 0 of 10 target points on the other fingers and made 0 of 6 pinches. With 22 it managed 1 and 3 simulationthe model’s thumb is weak, see chapter 6.

One paper pulls the other way. A hand with 6 driven joints scored 109 of 120 on six tasks against 97 of 120 for a hand with 20, though the two differed in hardware and in control policy paper.

About Fig. 1A flat model made for this page. Joints turn in proportion to moment arm ÷ stiffness, as in our simulations, and stop where a bone meets the part. The moment arms are the human values above; the springs are ours, chosen to show the two orders. A filled balloon marks a joint the part has stopped.

A weak grip looks like a shortage of force. Fit a bigger motor, pull harder, hold tighter. To see where that goes wrong, take the simplest finger there is.

One tendon is fixed at the fingertip and runs down the palm side of three joints to one motor. Pull it and the finger curls. That is how six motors can run a hand: one for each finger, and one more to swing the thumb across.

But the tendon pulls on all three joints at once. Which curls first? At each joint the tendon passes a small distance from the hinge, called the moment arm, and the pull times that distance turns the joint. In a human index finger the deep flexor tendon sits about 15.5 mm from the knuckle, 10.7 mm from the middle joint and 7.7 mm from the last one papersix hands. Against it stands whatever resists: a spring, an elastic hinge, a tight skin.

In our simulations the rule was plain. Each joint turns by its moment arm times the pull, divided by its stiffness, so the closing order is set by moment arm over stiffness and by nothing else simulationin five separate studies. Pull harder and every joint turns further, in the same order.

Begin with the soft base. The first bone swings out over the part, the second folds down behind it and the fingertip closes underneath. Let go, and the part settles a millimetre or two into the cage.

Now press the other button, which swaps the springs. The tip curls first, and the finger rolls into a fist before it has gone anywhere. When the base does turn, the fist comes down on top of the part. Let go, and the part falls. Push the slider to the end: joint 1 is stopped, the finger is as closed as it will ever be, and the part still falls.

In this model the pull sets how far the joints go, and the springs and moment arms set the order. A stronger motor reaches the wrong shape sooner. The order is decided when the finger is drawn, long before anyone chooses a motor.

Cross-talk sounds like the killer. One joint moves and eight tendons change length. Multiply by 22 joints and the hand seems to fight itself. So we took the worst place for it and studied that alone.

The worst place is the knuckle, where the finger joins the palm. It turns two ways, curling and swinging sideways, and every tendon bound for the two outer joints has to get through it. We modelled five designs with the same eight lines, over 441 poses simulationgeometry only, no forces.

In the naive design the passing lines cross the knuckle at a distance from its axes. Curl the knuckle and their path changes length by up to 18.87 mm. Hold the motors still and the outer joint turns 97° by itself.

The cure is geometry. Route each passing line so that it crosses each axis exactly on the axis line. A point on an axis does not move when the joint turns, so the path keeps its length. In our model the change falls to 0.38 mm and the drift to 0°. The 0.38 mm that remains is equal slack in a flexor and in its extensor, so the outer joint stays where it is.

So in this model cross-talk goes away when the part is drawn differently. The same table shows what does not go away. Look at the friction columns.

A tendon that bends round anything loses pull to friction, by a law much older than robots. The capstan law says that what comes out is what went in, divided by e to the power of friction times wrap angle. The size of the pulley does not appear in it. Only two things do: how far the tendon wraps, and how slippery the contact is.

A fist is three such bends in a row, 270° in all for the line that runs to the fingertip.

At a friction coefficient of 0.3, which we assumed for a bare line on printed plastic, the fingertip line in a fist loses 76 % of its pull. At 0.1 it loses 38 % and at 0.05 it loses 21 %. That is a formula with assumed friction. We have not measured a line in a printed channel. For scale, the friction of a human finger tendon at the middle joint has been estimated at 0.075 ± 0.021 paper51 hands, indirectly. Biology runs its tendons in a slippery sheath.

The cure for cross-talk also makes friction worse. To cross on the axis the lines must dive to it and come back, so the on-axis knuckle loses up to 59 % against 43 % for the naive one.

The other way out is to stop passing lines through the knuckle and put two small motors at the joint.

The motors cost torque and weight. Two micro gearmotors give 0.078 N·m continuously, which is 0.85 N at the fingertip, and put about 23 g at the joint. The tendon knuckle gives 0.22 N·m at 40 N of tension, 2.8 times as much, with 9.3 g at the joint, because its motors sit in the forearm simulationwith our arithmetic.

That suggests a hybrid finger: tendons for the two knuckle axes, where torque matters most and the line has passed the fewest bends, and a small motor or a fixed linkage for the two outer joints, where every extra bend costs the most. It is an idea, not a result.

If the motor stays, where does it go? It has two possible homes. One is the forearm, pulling tendons through the wrist the way your own body does. The other is the hand, at or beside each joint.

Figure tried the first and left it. Adcock says he chose forearm tendons in 2022, for space, for more degrees of freedom and because the body does it that way. The hand was built and tested in early 2023. His judgement three years later, and what he says of the next hand, are the four excerpts on the card above maker’s claimone post.

What that does not contain is a cause. None of the three sources of his that we read names a measured failure, whether friction, wear or stretch, or gives numbers for the tendon hand against any other our inferencefrom reading them.

1X ships the opposite choice. NEO’s hand drives 22 finger and palm joints with tendons from forearm motors, at low gear ratios of about 5:1 to 15:1. Its argument is feel, in the words on the second card. It reports up to 45 N of distal flexion force maker’s claim. It does not publish the tendon material, whether each tendon runs in a sheath, or a test of the whole hand our inferencefrom reading its page.

Tesla’s patent application sides with the forearm: an actuator there tensions a tendon that runs through the palm to a finger bone maker’s claimpublished 16 April 2026, not granted. A second application rearranges the cables across the wrist to limit their moment arms maker’s claim, a cure for a disease we meet in chapter 7.

On the other side, Wuji uses no tendons at all: “self-locking rotary direct drive joints”, four to a finger maker’s claim. Unitree’s Dex5-S has 22 motors in the hand maker’s claim. Sharpa’s Wave has 22 “Proprietary Actuation Modules” and does not say where they sit maker’s claim.

Papers do measure pieces of what a tendon costs. One finger delivered 33 N through a sheath 0.1 m long and 25 N through one 1 m long and curved differently paper. On a Shadow hand, a larger wrist pulley raised the mean time between tendon failures more than three times paper2019. Tendon trouble can be engineered down. Nobody says how far.

What would settle the argument is not exotic: two hands, one of each kind, at equal force and mass, on the same task, run until they fail, with friction loss, drift, re-tensioning and repair time logged. We found no such test in public our inferenceafter looking.

One of Adcock’s reasons for choosing tendons was “potentially higher degrees of freedom”. A forearm has room for many motors. The question is what each added joint costs. We built one model hand and gave it 6, 11, 16 and then 22 driven joints.

Count first. With six driven joints the model has 5 tendons and 40 guide points, the places where a tendon is held to a bone. With 22 it has 44 tendons, a flexor and an extensor for every joint, and 236 guide points simulationthese are route points in our model, not parts counted on a built hand.

What the joints buy is reach. The index fingertip of the six-joint hand can be put in 83 places on a 3 mm grid. With 22 joints it is 3,656. The thumb is the weak part of our model, and even at 22 it reaches only 1 of its 10 targets. It was never redesigned, so read that as a weak model and not as a law.

Table 4 · One model hand at four joint counts

Fingertip positions are counted on a 3 mm grid. A thumb touch is the thumb tip reaching one of ten target points on the other fingers and the palm.

What the joints cost is cross-talk. A tendon to the fingertip passes every joint on the way there, and when one of those joints bends, the line changes length. In the 22-joint model, turning one joint changes the length of up to 10 tendons for the thumb’s swing across the palm, 8 for each knuckle, 4 for each middle joint and 2 for each fingertip joint. If no joint disturbed another joint’s tendons, the table that links joints to tendons would have 44 entries. Ours has 125 simulation.

Now the word itself. Exponential means that each added joint multiplies the difficulty. Here is what can be counted.

• Motors, tendons and guide points grow in step with joints. From 11 to 22 driven joints our tendons double, and our guide points nearly do.
• Pairs of joints that could disturb each other grow with the square: 15 pairs at 6 joints and 231 at 22 (arithmetic).
• Friction in a tendon grows exponentially with how far the tendon bends. That is not the same as the joint count, and it is the subject of chapter 7.

None of these is a law of build effort, and build effort is what the claim is usually about. We looked for a family of hands with assembly hours measured at three or more joint counts and found none (our audit of the sources). The build times that exist come from single builders with no timed method: under 3 hours for MIDAS with 13 driven joints, about 4 for ISyHand with 18 motors, about 7 for RUKA with 11 and under 8 for ORCA with 17 papers [2] [3] [4]. They do not rise in step with the joint count.

What build logs do show is where the hours go. By our count of one build video, 76 % of the time spent putting a tendon finger together, 1,899 of 2,483 seconds, went on installing, tensioning and knotting tendons our inferencea step count.

A hand has hundreds of small parts, and somebody threads the tendons by hand. It is natural to think the price is labour. A parts bill says where the money goes before any labour is paid, and several research hands publish theirs.

In every bill that gives the split, the motors are the largest line, from 52.6 % to 81 %. The press reports the same shape for Chinese commercial hands: drive modules are more than half the cost, and the small coreless motors alone about 41 % press.

Now set the hours beside the bills. The slowest build in the table takes under eight hours. For labour to cost as much as the parts of ISyHand, its builder’s hour would have to cost $325 (our arithmetic: $1,300 over four hours). The hours are soft, but they are far from the range where labour leads.

The claim has a true core. A hand resists automation, which is the point of the 100 screws in chapter 1. That limits how many hands can be made. It is not where the published bills say the money goes.

A price is a different thing from a parts bill. It adds labour, testing, warranty and margin, and for hands on sale it varies widely.

Divided by driven joints the spread stays wide. It runs from about $24 of parts in the cheapest research hand, through $289 for the 22-joint hand with the lowest list price, to over $1,000 (our arithmetic on the numbers above). Touch is extra: tactile film is reported to add RMB 3,000 to 5,000 to a hand press.

About Fig. 10Tap a row to read what was measured and where it comes from. Filled marks are measurements. Hollow orange marks are claims or reports with no test method, and an arrow means “more than”. The grey band is one year at 450 to 1,200 cycles a day; the dashed line is one year at the rate you set.

A hand that costs $6,350 and works for five years is cheap. The same hand worn out in a season is not. Life decides the real price, and it is the number with the least behind it.

The claim comes from spec sheets. 1X says components and whole finger assemblies were tested for millions of cycles, and that the wrist joints “have been proven reliable well beyond 2 million cycles under high loads” maker’s claim. That is a statement about parts and about the wrist. It describes no test of the whole hand. Unitree lists “>100000 fatigue cycles” for one family of hands, with no conditions maker’s claim. Honda reports 450,000 cycles for its hand, and 24,000 at 5 kg maker’s claim.

Papers measure less and say more. The longest runs on research hands are a few thousand cycles, stopped by the clock and not by a failure. ORCA’s abstract reports more than 10,000 cycles and its methods a separate run of 2,250 grasps; the paper does not reconcile the two, and neither ran to failure paper.

The longest run to a real failure that we found is on a prosthetic hand: “494,000 cycles before the tendon breaks” paper. Printed parts fail much earlier: springs at 2,446 ± 499 cycles, soft actuators at 3,439 to 6,410 and a flexure finger at 45,000 papers [2] [3].

What fails first is often not the mechanism. In a soft finger with a camera inside, the picture went bad at 4,968 cycles while the mechanics were fine past 15,000 paper.

How many cycles are needed? One prosthetic hand has a battery sized for 1,200 open-and-close cycles a day maker’s claim. At 450 to 1,200 a day a year is about 165,000 to 440,000 loaded cycles, and no printed soft hand we found is within ten times of that our inference. A warehouse asks more: one press report gives the leading hands at one company 300,000 to 500,000 cycles, which it puts at about a month press.

We checked nine companies, and none publishes a loaded whole-hand life test or a failure rate (our audit of the sources). For scale, one industrial two-jaw gripper is rated at 5 million cycles maker’s claim. That is the kind of number hands will be held to.