Inside NEO's new hands: the tendon-driven, force-transparent design 1X just shipped

About a month ago I sat down for a podcast recording with Dar Sleeper, the head of product design at 1X, and he showed me video of the company's new humanoid robot hands that was basically mind-blowing: fast, precise motion I hadn't seen anywhere else. I mentioned it in my story but couldn't reveal exactly what I'd seen.

Today I can ... and NEO's hands are even more impressive than I thought at the time.

Check it out:

For a technical audience, this launch is worth pulling apart carefully, because 1X didn't just ship more degrees of freedom. It shipped an impressive new architecture — a hand designed as a sensor rather than just an actuator. 

Here's the engineering story.

The spec sheet

The headline number, of course, is 25 degrees of freedom: 22 fully actuated DoF in the fingers and palm, plus 3 at the wrist. Critically, 1X says all 25 are natively force-controlled and fully backdrivable ... not just position-controlled joints with sensing bolted on afterward.

The force and precision figures are where it gets interesting:

• Thumb CMC torque: 3.5 Nm
• Finger MCP torque: 2.6 Nm
• Distal flexion force: up to 45 N at the fingertip
• Wrist torque: 17.75 Nm
• Positioning accuracy: ±0.2 mm

Plus, there's one more really, really cool feature that turns out to be hugely important for a robot you might ask to make your kid a peanut-butter sandwich:

• Waterproofing: IP68, with food-safe materials throughout

That combination — high whole-hand force alongside sub-millimeter precision — is seriously impressive. With that, NEO can push a loaded cart or lift a ~20-pound load, then turn around and pick a grape off its stem or insert a light buld.

it can even pick up a small screw off the floor.

The big story: a "read-write" hand

Most robot hands are effectively "write-only." 

You command a position, the hand goes there, and no meaningful data comes back. The culprit is the transmission: at the 100:1 and 200:1 gear ratios common in the field, friction swallows contact forces before they ever reach the motor. The hand is essentially numb through its own joints, so builders wrap it in external sensors and infer what's happening at the fingertips.

That's why you have cameras pointed at hands that can't feel.

1X's approach is quasi-direct-drive. The hand runs proprietary tendons through what the company calls the 1X Tendon Drive at gear ratios of roughly 5:1 to 15:1, an order of magnitude lower than the norm. At those ratios the transmission is transparent enough that external forces backdrive the joints, and the motor reads them. Push on a finger and it yields ... and kind of reports exactly how hard you pushed. 

1X calls this force transparency: force flows out and information flows back along the same physical path, turning every push into a measurement.

(Note: this doesn't mean cameras in hands are necessarily bad: robots for warehouses or factories might need the most durable hands possible, and that might lead to different decisions about parts and processes.)

There's a second channel riding alongside it: proprioception. (That's the very humane ability to know where your hand is without looking.)

Because every joint is closed-loop, the hand always knows its own configuration without vision. That's pose plus effort, across all 25 joints, continuously. For a manipulation policy, that means every grasp arrives effectively pre-labeled with force, contact, and position data. 

Interestingly, that's the part that matters for training: the hand isn't just executing, it's generating supervised data on every interaction.

Where the motors actually live

A design detail worth calling out: the motors aren't in the hand. 

They sit in the forearm — the same place most of your own grip strength comes from — pulling tendons through the wrist. That's how a hand this light produces 45 N of distal flexion and 17.75 Nm at the wrist while staying cool enough for continuous operation. It also keeps distal inertia low, which, as we'll see, is central to how the hand stays safe.

Here's what that looks like:

This is deep vertical integration: in-house 1X Motors, custom electronics, embedded sensing, the proprietary tendon system, compact transmissions, and hand-specific firmware, all co-designed. 

1X's argument is that this tight coupling enables iteration speed and compounding gains that a modular, off-the-shelf approach can't match.

The skin: tactile sensing with shear

NEO's hand isn't hard fingers on a claw. 

There's a skin layer doing real work: high-resolution tactile sensing across the fingertips and surfaces that measures normal force, contact location, and — importantly — shear, the lateral force that signals an object is starting to slip. 1X essentially views tactile data as an image, with its own dynamic range, resolution, channels, and field of view.

The payoff is a reflex loop: feel any initial slipping through the shear channel, then re-grip through the transparent joints before the object is gone. That's faster than a vision system could catch it. 

1X is explicit that vision alone is insufficient for small, transparent, deformable, or occluded objects, which is exactly where contact feedback earns its keep.

The skin is co-designed with the sensors inside it and the tendons behind it ... a functional material, not a cosmetic one.

Durability, compliance, and the fact that it can wash itself

Pretty obviously, a hand that learns by touching everything has to survive touching everything. Durability is a huge challenge for robot hands. It's relatively easy to make a great hand that works a few times; it's much harder to make a great hand that works for years and years.

(Yeah, "relatively" is doing a lot of work in that previous sentence.)

1X says its components and finger assemblies have run through millions of test cycles. The wrist joints are proven well beyond 2 million cycles under high loads, 1X says, and drive units have been tested at temperature extremes.

Compliance is where the low gear ratios pay a safety dividend. Because the ratios are so low, the tendon drive so transparent, and distal inertia so low, external impacts simply backdrive the fingers instead of resisting them. 1X's slow-motion footage shows the hands yielding when slapped, hit with a hammer, or pinched in a closing drawer. 

For a machine meant to work around people, that's compliance by construction rather than by software.

It's also additional durability because the hand is not so rigid that impacts instantly damage it.

And the detail that I love: the whole hand is sealed to IP68 with food-safe materials, so NEO can wash its own hands at the sink. Think about what that enables: if NEO fries you an egg or chops vegetables for a salad, it can clean up afterward instead of spreading the mess around your kitchen. 

I'm not aware of any other humanoid that can do this today.

The strategic number

Buried at the end of 1X's announcement is the number that actually matters: hundreds of these hands have already come off a scalable production line, and the company says it has capacity to build 10,000 hands this year: every unit made in-house end to end, from tendon materials and 1X motors to the soft polymers, skin, and tactile stack, with full end-of-line testing.

Again, that's huge.

A hand that can't be built at scale can't run experiments at scale, and without contact data at scale there's no path to general-purpose manipulation. As 1X puts it, every grasp is an experiment ... and they're about to run a lot of them.

Why hands are the whole ballgame

1X frames a humanoid as "a computer whose API is its hands," and I think that's exactly right. 

The legs get the robot somewhere; the body houses the battery and compute; the hands do the actual work. A two-finger gripper exposes essentially three verbs to developers — pick, place, push — and every application is a blind composition of those three, forever. 

There's real runway here, too. 

When I spoke with Sleeper a month ago, he told me NEO will ship "with more hardware capability than the AI currently knows how to use." In other words, the hardware is ahead of the software, so over-the-air updates should make each NEO progressively more capable over time as the models catch up to what the hardware can already do.

Great hands are the prerequisite for great humanoid robots, and there's a real race on to build them. 

Genesis AI recently emerged from stealth to unveil the hands behind its Eno humanoid, and Kyber Labs is building impressive hands as a dedicated component supplier rather than a full-robot company. I keep coming back to something Geordie Rose, co-founder of Sanctuary AI in Vancouver, once told me: robot hands are half the complexity of a robot. Look at the motor, gear, and tendon count in a top-end tendon-driven hand and you can see why.

They need to be strong, flexible, cleanable, durable, and about as capable as our own ... an almost impossibly high bar. But seeing what 1X just shipped, we're closer than I expected to be at this stage.

NEO is aimed at the home for now. Assuming 1X's AI is pretty awesome (and yes, that's a big assumption) I'm more excited now about a NEO in my home than I was before.