The hand is the hard part

Moravec's paradox, formulated in the 1980s, says that the things humans find hard are easy for computers and the things humans find trivial are nearly impossible. Chess fell in 1997. Language fell around 2022. Picking up a crumpled sock has not fallen. The hand is where the paradox lives.

What a hand has to do

Grasping is not one skill. A person switches between a pinch, a hook, a power wrap, a tripod and a lateral key grip dozens of times an hour without thinking about it, and regrips mid-motion when the object shifts. Doing this requires three things at once: enough joints to form the shapes, enough sensing to know what is happening at the contact point, and a control loop fast enough to react before the object slips.

Robots are good at the first, poor at the second and improving at the third. Building a hand with twenty joints is an engineering exercise. Building one that knows it is holding a wet glass is not.

Why touch is so difficult

A camera sees an object until the hand covers it. At the exact moment grasping becomes delicate, vision goes blind and touch has to take over. Tactile sensing has to survive being squeezed thousands of times a day, work through a skin that is thin enough to feel but tough enough to last, and produce a signal a controller can use in milliseconds.

Three approaches compete. Vision-based tactile sensors put a small camera behind a soft gel pad and watch the pad deform, which gives extraordinary spatial detail and costs a camera per fingertip. Capacitive and resistive arrays are cheap and thin but coarse. Force and torque sensing at the joint infers contact from motor current, which needs no fingertip hardware at all and cannot tell you where on the finger the contact is.

Most shipping hands use the third, because it survives. Most research results use the first, because it works.

The durability problem nobody films

A production hand has to complete millions of cycles. Tendon-driven designs, which pull cables through the fingers from motors in the forearm, give a light hand and a strong grip, and they wear at the tendon. Direct-drive designs put a small motor in each joint, which is robust and makes the fingers thick and heavy. Gear backlash, cable stretch and dust all arrive in month four of a pilot, not in week one.

This is why the repair story matters more than the specification sheet. A hand designed as a replaceable module with a documented price is a product. A hand that requires the manufacturer's engineer is a prototype with a service contract.

What learning changed

The traditional approach analysed the object, computed a grasp and executed it. It worked in a lab and fell apart on anything deformable, transparent or unexpected. Learned policies changed the order: the robot tries, fails, adjusts, and the adjustment is what is learned. Trained across simulation and human demonstrations, these policies now generalise to objects the robot has never seen, which is the thing the analytical approach could never do.

The gap that remains is the long tail. Ninety percent success on a bin of mixed parts sounds impressive and means one failure every ten picks. At a thousand picks a shift, that is a hundred interventions. Useful automation needs somewhere above 99.5 percent, and every additional nine costs more than the last.

The alternative nobody likes to say out loud

For a specific, known set of objects, a custom gripper beats a general hand on every metric: cost, speed, reliability, maintenance. A suction cup moves boxes better than fingers do. A two-jaw parallel gripper handles machined parts perfectly. The entire case for the five-fingered hand rests on the value of not having to choose in advance.

That case is real in a job shop with a thousand part numbers and weak in a plant that makes one product. Which is why the first hands that earn their cost will probably appear in mixed, low-volume work: hospital logistics, laboratory sample handling, retail backrooms, repair.

What to watch

Watch for published cycle-life numbers on a fingertip sensor. Watch for a hand sold separately from a robot, which would mean a component market exists. And watch grasp success rates measured on a bin that the company did not curate.