Why humanoid robot hands are the hardest part. Offbrand Human

A robot arm needs only about six joints to reach any point in a room from any angle. One human-like hand needs more than twenty. That imbalance, a whole arm cheaper than a single hand, is why manipulation is the wall the entire industry keeps hitting, and why the humanoid robot hand is the first row to look at on a spec sheet.

The six-versus-twenty problem

Start with the arm, because the arm is the easy part. Reaching any pose (a position and an orientation) in three-dimensional space takes exactly six degrees of freedom, six independently driven joints. Most robot arms use six or seven. Seven gives you a little redundancy, the same trick your own arm uses to reach around a box without moving the box.

Now the hand. Getting there is the easy half. The hand also has to conform to whatever it finds when it arrives: a mug handle, a zipper, a grape it must not crush. That means many small joints, packed into the volume of a human hand, all controlled at once. The numbers below are the actual bill of materials, and they are startling next to "six joints for a whole arm."

The spec ladder

A ladder of degrees of freedom, fewest first, summarizing the spec table below it: a whole robot arm needs only 6 to 7 joints to reach any point in a room at any angle, Tesla's Optimus hand had 11 in 2023, the Shadow Dexterous Hand has 20 across 24 joints with 20 motors up in the forearm pulling tendons and 34 tactile pads per fingertip, Tesla's Gen 3 Optimus hand doubled to 22 in 2024, and the 1X NEO hand claims 25 in 2026. Photographs sit beside four of the rungs: a complete Universal Robots UR20 industrial arm on the first, the black five-fingered Shadow Dexterous Hand on the Shadow rung, a single raised Optimus hand on the Gen 3 rung, and NEO's open mechanical hand on the top rung. The Gen 2 rung reads 2023, no close-up, instead of a photograph, because Tesla published none of that generation's hand.

Component	Degrees of freedom	The detail that costs money	As of
A whole robot arm	6 to 7	Enough to reach any pose in the room	baseline
Tesla Optimus hand (Gen 2)	11 DOF	The 2023 hand	2023
Tesla Optimus hand (Gen 3)	22 DOF	Doubled the joint count in one generation	2024
1X NEO hand	25 DOF	Marketed as closely replicating a human hand	2026
Shadow Dexterous Hand	20 DOF across 24 joints	20 DC motors in the forearm, Hall-effect sensor in every joint, 34 tactile pads per fingertip	research benchmark

The Shadow Hand row is the only one that spells out what those degrees of freedom actually require: twenty motors (they do not even fit in the hand, they live up in the forearm and pull tendons), a position sensor at every single joint, and touch sensors dotting each fingertip. Twenty motors, twenty-four joints, per-joint sensing, all miniaturized into something the size of your hand. That is why a serious dexterous hand can cost as much as the entire rest of the robot. It is a second robot bolted to the end of the arm.

The table's 11 DOF row in motion: Tesla's Gen 2 reveal shows the hand's party piece, fingertip pressure sensing demonstrated on an egg.

Why it is hard, in kitchen terms

Four problems stack on top of each other, and the hand is the only part where all four hit at once:

Cramming. Twenty-plus joints have to fit in a hand-sized volume. There is nowhere to put big, strong motors, so the motors move to the forearm and yank on tendons, which adds its own slop and control headaches. Reaching across a room, by contrast, has all the room in the world for big joints.
Feeling. To hold a grape without pulping it, the hand needs touch. Good hands sense force per finger (the Shadow Hand reads forces through those 34 fingertip pads). An arm reaching an empty pose feels nothing and needs to feel nothing.
Coordinating. All those joints must move together, in real time, to close around an unknown shape. Balancing a whole body is a coordination problem too, but the hand adds a second one at a much finer scale.
Adapting. Every object is a little different, so the grasp cannot be a fixed script; it has to adjust on contact, which leans on AI and learning rather than coming for free.

An arm solves one problem (get there). A hand solves four (get there, fit, feel, and adapt), simultaneously, in a tiny space. That is the whole reason walking demos are everywhere and reliable grasping is not.

1X's first-generation NEO standing full-length on a white studio sweep, a hooded gray knit suit covering the whole machine except the segmented white and black mechanical hands hanging open at its sides, which is the hand-sized volume every one of those joints has to fit inside. — The 2023 NEO on 1X's own product page: soft knit and padded feet everywhere, then bare rigid fingers at the cuffs. The hand is the one part they could not wrap. Photo: 1X

The people building them say the same

Melonee Wise, former chief product officer at Agility Robotics, said plainly that "currently AI is not robust enough to meet the requirements of the market" (IEEE Spectrum). She is talking about exactly this: the legs are close to solved, the hands are not. A smarter language-model brain does not fix a hand that fumbles a cup, because the bottleneck is mechanical and sensory as much as it is software. Manipulation stays the last thing to work, on every one of these machines.

The people selling the hands concede the premise too, then claim they have beaten it. Announcing NEO's hands in July 2026, 1X chief executive Bernt Bornich called them "nearing or surpassing human-level dexterity, strength, speed, and reliability". That post is where the 25 DOF row in the table comes from, and it is a promise with a name on it: the number is checkable on a delivered robot, and so is the word "reliability".

1X chief executive Bernt Bornich introduces NEO's 25 degrees of freedom, tendon driven hands, and says robotics spent seventy years working around the hand problem (@BerntBornich, 2026-07-09)

The maker's own source for the 25 DOF figure in the table above, and the dexterity claim this article says to hold against a shipped robot rather than a launch video.

What this means when you read a spec sheet

Two habits, both free:

Finger DOF and tactile sensing are the real "can it manipulate things" spec. A "25-DOF hand with tactile fingertips" is a very different (and far more expensive, and far less proven) machine than one with a simple three-finger gripper. A gripper is cheaper and more reliable because it does less. If a robot's job is picking up bins, that is the right answer, not a downgrade.
Treat high finger-DOF claims as the most expensive and least-settled number on the sheet. A robot can walk beautifully and still not close its hand around your coffee mug reliably. When you see an impressive full-body demo, the question that separates real capability from a good edit is always: what did the hands actually do, and did they do it twice.

How the joints themselves are built (and why the gearbox inside each one costs so much) deserves an explainer of its own. For how much of any of this the robot does on its own versus with a human driving, see How autonomous are humanoid robots, really?. And for which humanoids you can actually put hands on, the tracker lists every real one.