1X NEO Hands: 25-DoF Tendon Drive, Force Transparency, and 10K Units in 2026
1X shipped 25-DoF tendon-driven hands for NEO July 9, 2026 — force-transparent joints, tactile skin, IP68, ±0.2mm accuracy, 10,000 hands/year. vs Figure, grippers, and the data-for-embodied-AI bet explained.
On July 9, 2026, 1X Technologies published NEO's Hands | An API to the Physical World — not a gripper refresh, but a 25-DoF tendon-driven hand 1X will put on every NEO humanoid. CEO Bernt Børnich's thesis: match or beat human hands on dexterity, strength, safety, and reliability so data — not hardware — becomes the binding constraint on what humanoids can learn.
1X's headline framing is deliberately developer-centric:
A humanoid with a two-finger gripper exposes three verbs to developers: pick, place, push. Every application anyone writes on that platform is a composition of those three verbs, forever — executed blind.
Legs and perception get the robot to the world. Hands determine what it can do there and what it can know while doing it. That maps cleanly onto Yann LeCun's physical-agents argument: language models reason about text; manipulation requires closed-loop contact intelligence the way children learn — probe, feel, update.
Mistral Robostral Navigate showed how far map-less locomotion got in 2026. HN's instant reply to dexterity news: navigation ≠ pick up the arbitrary thing.Hands are the missing action space — exactly what 1X is shipping hardware for.
World-picture — five layers 1X stacks
1X describes dexterity as assembling a world-picture subsystem by subsystem:
Layer
What it adds
Without it
Rigid gripper
Blind position control
Dark patch — act, barely perceive
+ Force transparency
Contact returns through joints
No dynamic range in touch
+ Fine motion
Pose, trace, pinch, in-hand rotate
No small-object regime
+ Tactile & shear skin
Pressure, slip, contact location
Vision-only guesses on transparent/deformable objects
+ Reflex & robustness
Re-grip slips; survive millions of touches
Too fragile to learn by probing
The July post is 1X claiming all five layers ship together — not a lab prototype gripper swapped in for demos.
Force transparency — joints that read, not just write
Most industrial hands are write-only: command angle, hope contact matched intent, wrap external F/T sensors or cameras because 100:1+ gearboxes eat contact forces before they reach motors.
NEO hands use quasi-direct-drive tendons at ~5:1–15:1. All 25 DoF are natively force-controlled and backdrivable — push a finger, it yields, and reports effort. 1X calls this force transparency: the same path that applies force returns measurement.
Proprioception rides along: closed-loop joints always know configuration (eyes-closed fingertip touch test for robots).
Motors live in the forearm — human-analog — pulling proprietary tendons through the wrist so the hand stays light but can deliver multi-Nm grasps without overheating.
Those specs target the small-object regime where most household labor actually lives — screws, chargers, grapes, not pallet boxes.
Tactile skin — the last half-millimeter
Vision fails on transparent, deformable, occluded objects. 1X co-designed functional skin with embedded sensors measuring normal force, contact location, and shear — slip detection fast enough to re-grip before objects fall.
The post shows pressure heatmaps on handshakes and origami grasps without crushing — evidence the control stack closes loop on contact, not just kinematics.
Safety by construction — backdrive, not brute lockout
1X pairs low gear ratios and low distal inertia with compliance: slow-motion tests show fingers yielding to slaps, hammers, closing drawers, and foam impacts. A hand that learns by touching everything must be gentle by physics — relevant anywhere humans and robots share space.
Hands are IP68 and food-safe — NEO can wash its own hands at a sink. That is a product requirement for home deployment, not a lab flex.
10,000 hands — why manufacturing is the strategy
The strategic number in the post is not peak torque — it is 10,000 hands in 2026:
A hand that can't be built at scale can't run experiments at scale, and without data at scale there is no embodied AGI.
1X claims hundreds already produced on a dedicated line with in-house motors, electronics, tendons, skin, and tactile stack. That contrasts with Figure AI's robot-outnumber-humans factory ramp — different scale story (whole humanoids per hour vs standardized hand modules), same data flywheel logic.
Every grasp becomes a labeled experiment: joint forces, tactile images, poses. That is the dataset Indian egocentric camera workers are manually collecting today — 1X wants the hand to generate rich labels natively during deployment.
June 2026 context: 1X launched its World Model Lab (Sam Sinha, head of world models) — pairing hardware that acts with models that predict physical consequences, adjacent to NVIDIA Cosmos 3 physical AI and world-model research on explainx.ai.
1X NEO hands vs Figure and grippers — honest comparison
No independent benchmark yet compares 1X vs Figure on identical manipulation suites — treat demo reels as directional, not scored. The meaningful claim is hardware ceiling removal, not "NEO beats humans on every task today."
What to watch next
Third-party evals — LEGO and USB-C demos are marketing-grade until independent labs score success rates
Developer SDK — what APIs expose force, tactile, and proprioception streams to policies
World Model Lab outputs — do 1X world models train on hand-native contact logs?
