1X Built Robot Hands That Pick Grapes and Speak Sign Language

Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters.

1X gave its NEO humanoid a new pair of hands, and the company is making a claim far bigger than the hardware itself. It says the physical limits on what a robot can do with its fingers are finished, and everything that happens next comes down to the software learning to use them.

Let's get into it.

On 9 July 2026 the Norwegian-American company 1X, founded in 2014, revealed a redesigned hand for NEO that carries 25 degrees of freedom, with 22 of them in the fingers and palm and three more at the wrist.

The hands ship on every NEO robot rather than sitting in a lab as a demo, and 1X frames them as the piece that decides whether a humanoid is useful at all, since a robot that walks across a room and cannot handle what it finds there is closer to furniture than help.

CEO Bernt Børnich put it plainly when he said the robot can now do the things people do with their hands every day, and called it the moment the industry has been waiting for.

How the Hands Feel

The mechanical trick sits in how the fingers move. Instead of cramming motors into the hand, 1X placed them in the forearm and ran cables down through the wrist to pull the fingers, a tendon-drive layout that keeps the hand light while still delivering a firm grip.

What makes it unusual is the gearing, which runs at low ratios of roughly five to one up to fifteen to one, far below the heavy gearing most robot hands rely on. That keeps the fingers backdrivable, so pushing on one makes it yield while it reports back exactly how hard it was pushed.

1X calls this force transparency, and the effect is that every joint acts as both a motor and a sensor at once. High-resolution tactile sensors across the fingertips and palm add another layer, reading pressure, contact location and the sideways shear that tells the hand a glass is starting to slip before it falls.

What NEO Can Do With Them

The demonstration reel is where the range shows. NEO assembles LEGO models, picks individual screws and coins out of a wallet, spins light bulbs into sockets, uses a screwdriver, rotates objects inside a single hand, zips up a jacket, sorts grapes by colour without bruising them, pours tea from a kettle, catches a soft ball, plugs in a USB-C cable, holds a wine glass and even signs in sign language.

One outlet that got an early look described the fingers as freaky fast, and the speed is real, though the more important quality is subtler, the ability to feel what it is holding and correct before something drops. The strength numbers sit behind the delicacy, with the thumb generating peak torque of three-point-five newton metres, the wrist reaching seventeen-point-seven-five, fingertip flexion pushing up to forty-five newtons and positioning accuracy rated at two-tenths of a millimetre.

The hands are sealed to IP68, built from food-safe materials and washable, so NEO can rinse them under a tap after handling food rather than needing a technician.

The Part the Demos Leave Out

The footage is convincing enough that it is worth reading the fine print. NEO is not fully autonomous, and 1X has confirmed that not every viral clip ran on its own. For anything complex the robot leans on an Expert Mode, where trained human operators connect remotely through its cameras and guide it through the task.

The company says operators join only when a user asks for help and that faces and sensitive details can be blurred, but the distinction matters, because a hand that can physically pick a grape is a different thing from a robot that decides on its own how to pick it. The hardware in these videos is real, and the intelligence steering it is often a person sitting somewhere else.

Why 1X Built a Factory for Hands

The more telling number is not on the spec sheet at all. 1X has already produced hundreds of these hands and says it can build ten thousand in 2026, on a line that makes the motors, tendons, polymers, skin and tactile sensors entirely in-house.

The hardware has been run through millions of operating cycles in testing, with the wrist joints alone surviving more than two million load cycles before shipping. That vertical stack is the strategy rather than a footnote, since a hand that cannot be manufactured at scale cannot generate the volume of real-world practice a robot needs to get better.

The competition is running the same race, with rival AGILINK having already shipped eight thousand hands, which turns the humanoid contest into a supply-chain and training-data problem as much as an engineering one.

The Bottleneck Moves to the AI

Underneath the demos is a genuine shift in where the hard problem lives. For most of robotics history the hand was the wall, too clumsy or too numb to do useful work, so engineers built systems that worked around it. 1X is arguing that wall is gone, and that the limit on what NEO can do now sits in the AI and the training data rather than the fingers.

That reframes the whole effort, because if the hardware can already match a human hand on fine tasks, then progress becomes a question of how fast the software can learn to drive it, and how much practice the company can feed it. It is the same logic that pushed 1X to treat manufacturing as the real moat, since without hands built at scale there is no data at scale, and without data there is no path to a robot that acts on its own.

The Bottom Line

1X has built the most capable robot hand the industry has shown, and it has done the harder thing of building it on a line that can turn out ten thousand a year. The honest caveat is that a person is still steering NEO through the trickiest moments, so the grape-picking clips say more about the hardware than about the robot's mind.

But the direction is clear, and if 1X is right that the bottleneck has moved from the hand to the software, then the next race is not about better fingers at all, it is about who teaches a robot to use them first.

Category: Automation Planning

"Act as an automation strategist. I run [describe your team or business] and I want to work out which tasks are ready to hand to a machine or AI and which still need a person. Here is a list of what we do day to day [paste your tasks]. Sort them into three buckets, the ones safe to automate now, the ones that need a human in the loop for judgement or safety, and the ones not worth automating yet, and for each bucket explain the one factor that put it there so I can reassess as the tools improve."

There is a pattern worth noticing here that reaches well past robots. The hard part of almost any capable system tends to migrate, and once the obvious limit falls the real work moves somewhere less visible, from the hand to the brain, from the tool to the judgement about when to use it.

Watching where the bottleneck goes next is usually a better guide to what matters than watching the part everyone is filming.

Hit reply, I read every response.

———————————————————————————

See you in the next one.

— Vivek

P.S. Know a founder or engineer betting on humanoid robots reaching the home? Forward this to them. They can subscribe at https://savvymonk.beehiiv.com/