1X NEO's 25-DoF tendon-driven dexterous hands

The headline number is 25 degrees of freedom, but the engineering story is buried in a ratio most spec sheets never print. Conventional robot hands drive their joints through gearboxes of roughly 100:1 to 200:1, which multiplies motor torque enormously but at a cost: the joint becomes numb. It cannot feel back through all that gearing, and it holds position rigidly regardless of what it bumps into. 1X went the opposite direction, running quasi-direct-drive tendons through its 1X Tendon Drive at gear ratios of only about 5:1 to 15:1 ^[1]. That single choice is what makes each of the hand's joints behave as both a motor and a force sensor - the property engineers call backdrivability or force transparency.

The practical consequences are what the demos are really selling. Because so little gearing sits between the motor and the fingertip, the hand can measure force directly at every joint and yield when it meets resistance - it gives when hit by a hammer or caught in a drawer instead of crushing through. 1X layers on top of that high-resolution tactile sensing that reads normal force, contact location, and shear across the fingertips, which is what lets the hand detect a slip and re-grip in real time ^[3]. Analyst John Koetsier, who covered the launch, framed exactly this - backdrivability and force transparency - as the differentiator that separates these hands from the stiff, numb grippers that have dominated humanoid demos, arguing that hands, not locomotion, are the gating component for a genuinely useful home robot ^[2].

The reveal is unusually dense with hard specifications, and the numbers tell the story better than the adjectives do. The hand carries 25 total degrees of freedom - 22 actuated across the fingers and palm plus 3 at the wrist - all force-controlled and backdrivable, with no passive spring-loaded joints ^[1]. The tendon drive runs at gear ratios of roughly 5:1 to 15:1, against the 100:1 to 200:1 of conventional designs ^[2].

Strength and precision sit side by side. Peak torque reaches 3.5 Nm at the thumb CMC joint and 2.6 Nm at the finger MCP joints, with distal flexion forces up to 45 N and wrist torque of 17.75 Nm - enough to lift a roughly 20-pound kettlebell while still picking grapes without bruising them ^[2]. Positioning accuracy is quoted at plus or minus 0.2 mm. On durability, 1X says the wrist joints have been validated beyond 2 million cycles under load ^[3]. And the number that reframes the whole launch: an in-house production line built to turn out 10,000 hands in 2026 ^[1].

The most revealing line at the launch was not about torque. Dar Sleeper, 1X's head of product design, said NEO will ship with more hardware capability than the AI currently knows how to use ^[2]. Read plainly, that is a deliberate strategy: build a hand that can already do more than the models can drive, so the limiting factor on what NEO can accomplish is AI and training data, not physical capability. 1X reinforces the framing by calling the hand the robot's API to the physical world, treating dexterity as the software-like interface layer for household tasks ^[1].

This is where the 10,000-hands-per-year line stops being a manufacturing brag and becomes the actual plan. 1X builds the motors, tendons, electronics, tactile sensors, and soft polymer skin itself, and it argues that this vertical integration is a moat rivals cannot close simply by sourcing better components ^[4]. Shipping hands at volume is not only about revenue - every deployed hand is a data-collection endpoint, feeding the real-world manipulation data the models need to catch up to the hardware. The bet is circular by design: over-build the hardware, ship enough of it to gather the interaction data, and let that data close the software gap the hardware deliberately opened.

The response online split along a clear seam. Accelerationist communities greeted the reveal with near-unanimous excitement, seizing on the safety story - a hand that yields when caught rather than crushing - and on the leap in tactile capability, with the dominant framing casting physical manipulation as the next abstraction layer above software Computer Use. But the sharpest voices raised a harder question the marketing does not answer: how much of the polished demo choreography is autonomous, and how much is a human driving it. 1X's own product includes an Expert Mode remote human-assist path, and skeptics argued the trailer was likely at least partly teleoperated - a distinction that decides whether these hands are a robotics milestone or a very good puppet.

The grounded caution runs deeper than the autonomy debate. Independent reporting notes there is no third-party verification of tendon cycle life, real-world failure modes, or how the hands hold up outside a controlled demo ^[5]. One commenter with hands-on grape-destemming experience argued the precision on display, impressive as it is, is not yet good enough for precision-critical professional work. That gap - between a compelling home-chore reel and durable, autonomous, professional-grade reliability - is exactly what independent testing over the coming year will have to settle, and it is the reason the 2-million-cycle and 10,000-unit figures matter more as claims to be checked than as settled facts ^[3].