Anyone who has ever worn a cast on their hand knows how quickly daily life gets complicated. Pouring water, for instance. Grip the bottle, feel the resistance, adjust the force while the weight shifts with every degree of tilt. A healthy hand works all that out in passing, every second, without us noticing any of it.
With a cast on, you stand in front of a glass of water as if it were an engineering problem. That engineering problem is exactly what a whole industry is currently wrestling with.
27 bones, more than 30 muscles, over 20 degrees of freedom, grip forces of more than 400 newtons, all at a weight of roughly 400 to 500 grams. The real trick does not even sit in the hand. The strongest muscles are in the forearm and pull through tendons, which keeps the hand itself light and fast. And in the palm, around 17,000 touch fibres report at every moment whether something is slipping, giving way or about to break.
The manufacturers' human-form argument says our world is built for the human body. For reach, gripping height and two arms that hold and join at the same time, that is true. For legs there is a ready-made alternative out of a hundred years of industrial history. The wheel. For the hand there is none. If the legs are a bet you can argue about, the hand is the bet that has to pay off. And of all of them, it is the least solved.
Parallel grippers and suction cups have automated structured tasks with rigid objects for decades, and they do it superbly. Where they fail is at what makes human work human. Deformable objects, tool changes, anything where contact is the point. Plugging in a cable, closing a zip, cracking an egg. Anyone who wants the universal robot cannot get past a universal hand.
Why this is so hard comes down to a sentence the industry says about itself. Robotics has no internet. For language and images, bodies of data have grown over decades. For what happens during a grasp, there is nothing comparable. Force, give, slip, none of that is in any dataset, because nobody ever wrote it down. And if you map a person's movements onto a two-finger gripper, most of it is lost on the way, because a gripper takes hold differently than a hand does.
How serious this data problem is shows up in a side theatre. Around suppliers like MANUS, a separate industry has grown up building specialist gloves to capture what human hands do. Camera-based tracking fails precisely in the moments that matter, when fingers overlap or a grip hides the fingertips. So the gloves measure on the body rather than from outside, and people use them to teleoperate robot hands in order to generate training data. A lot of what looks autonomous in demos is exactly that. A human in gloves.
How far that effort goes is clear from an ETH spin-out in Zurich. In mid-July, alongside its own robot hand, mimic robotics presented a second device, an exoskeleton of rigid links rather than a glove. It only lets the wearer's hand make the movements the robot hand can execute, with touch sensors, joint encoders and a camera at the wrist, all mapped one to one onto the machine.
The second answer to the same gap is simulation. Instead of collecting data, you generate it. In environments like NVIDIA's Isaac Sim, thousands of virtual robots practise at once and produce more attempts overnight than a real machine manages in years. For walking, this works remarkably well.
The catch has a name of its own, the sim-to-real gap. Every simulation is an assumption about how material behaves, and the closer you get to contact, the worse the assumption gets. How far a screw slips in a fingertip, how a seal gives, how a cable settles, these are among the hardest quantities to compute at all. When walking, a controller absorbs errors like that and the robot recovers. When gripping there is nothing to correct. The part is either held or it is on the floor.
Both routes lead back to the same place. You can film a movement and you can compute it. What happens between fingertip and workpiece escapes both.
An ideal robot hand has to be dexterous, hold reliably and stay affordable, and research has shown for years that every existing design sacrifices one of those corners. Prioritise precision and you put the motors straight into the joints, which makes the hand large and expensive, and the gearboxes overheat under sustained load. Build compact and human-sized and you go tendon-driven, and those tendons stretch, fray and snap. The architecture that produces the dexterity is the same one that wears out. Push the cost down and you cut degrees of freedom and tactile resolution. The industrial gripper escaped this triangle decades ago by giving up dexterity. Two fingers, but five million cycles, 24 months of warranty and affordable.
Anyone who thinks this field belongs to the startups alone is missing who is entering it and with what. Schunk of Lauffen am Neckar is the world market leader in gripping and clamping technology, a third-generation family business. By its own account it has been researching human-like hands for around twenty years, long before there was a market for them. In January 2026 that became a company of its own aimed at series production, and since June it has been working with Bosch. A prototype has been announced. A cycle count of the kind that appears on the gripper datasheets from the same company has not been communicated so far.
mimic in Zurich presented a hand in mid-July that is explicitly built for industrial operation, and justifies its construction almost entirely through wear. The motors sit in the forearm and pull the fingers through tendons, the way a human does. Instead of running the tendons through guide sheaths, they route every single one over bearings and rollers, because sliding friction eats material and rolling friction eats far less. They revised the number of joints downwards based on data from real industrial tasks, dropping every joint with a small contribution in favour of service life. The datasheet reads like a supplier's, with back-drive torque, backlash and continuous load.
1X does quote a cycle count. More than two million cycles under load, plus IP68 protection and food safety. The catch is in the small print. The two million apply to the wrist. For the tendon-driven fingers it only says that assemblies have run through millions of test cycles, with no number. Precisely where the wear question is decided, it stays vague. And all of it is manufacturer data, not verified field values.
Sharpa optimises in the other direction, for feel. More than a thousand tactile points per fingertip, driven through fine gearing rather than tendons, awarded at CES. Sharpa quotes a number too, more than a million grip cycles, again self-reported. How that holds up in shift work nobody outside the company knows yet. At NVIDIA GTC in the spring the hand was teleoperated through data gloves in a much-watched demo, including catching a thrown ball.
According to estimates from the supplier industry, the hands account for between 15 and 25 per cent of a humanoid's bill of materials, and more than 30 per cent where there are many degrees of freedom. Unlike an industrial robot, they are mandatory equipment, two per machine. Demand scales one to one with unit numbers, and so does the cost.
What that means in practice shows up at the manufacturer with the highest volumes. Unitree's cheapest humanoid costs 4,900 dollars and comes with rigid fists that cannot grip anything. For that model, hands cannot even be ordered. They start with the developer model at 10,500 dollars, and there only as an option. Even the flagship carries dummies in the base version, which complete the silhouette and nothing else. How many machines finally ship with working hands is something the company does not publish.
That is the real answer. The most expensive and at the same time most failure-prone part sits exactly where the product promise is decided, and the volume leader leaves it off when in doubt. What remains is an expensive carrier for a gripper that has not been solved yet.