How 1X Built NEO’s 25-DOF Tendon-Driven Hands

Walking brings a humanoid within reach of a task. Household work then depends on whether the hand can grasp objects with different shapes, stiffness, surface friction and fragility without dropping or crushing them.

1X's July 2026 engineering release describes NEO's integrated hand and wrist as a 25-DoF tendon system with 22 fully actuated finger and palm axes plus three wrist axes.

Most objects inside a home were designed around the dimensions and movement of a human hand. Door handles, glasses, dishes, clothing, electrical plugs and cleaning tools all assume flexible fingers and continuous force adjustment.

A basic parallel gripper can move boxes or repeat a fixed factory task. It struggles when an object changes position, deforms under pressure or requires several contact points at once.

Household work creates constant variation. A robot may need to hold a plate firmly, touch a glass gently and reposition a piece of clothing without a predefined grasp. That makes hand design one of the main hardware limits for domestic humanoid robots.

1X says the 25-DoF count includes 22 fully actuated finger and palm axes plus three wrist axes. Tendons route force from actuators through the hand while keeping more motor mass away from the fingertips.

The company reports low-ratio transmissions of roughly 5:1 to 15:1, force control and backdrivability at every joint. Those choices support contact control, but tendon tension, routing friction and wear still require calibration and maintenance.

1X reports peak distal flexion force up to 45 N and wrist torque of 17.75 Nm. These figures should remain attributed to the company until comparable independent test data are available.

Finger position alone does not tell a robot whether an object is secure. The hand also needs information about pressure, contact location and movement between the object and the fingers.

Tactile sensing can help the controller detect when a grasp is too weak, when an object begins to slip or when additional force could damage it.

This feedback is especially important for transparent, reflective or deformable objects that remain difficult for vision systems. A camera may identify a cup while tactile sensors reveal whether the fingers are actually holding it.

The useful measure will be how consistently the system combines vision, touch and force control during long sequences of household manipulation.

Industrial robots often rely on rigid structures, physical separation and fixed workspaces. A home robot cannot assume that people will remain outside a safety fence.

NEO uses compliant mechanics to reduce the force created during unexpected contact. This does not remove the need for perception, motion limits and emergency stopping. It gives the hardware another layer of protection when software reacts too slowly or an object moves unexpectedly.

Quiet operation also matters inside a home. A robot that works near people for several hours cannot produce the noise, speed or abrupt movement accepted inside some industrial environments.

The hand therefore has to balance force, speed, durability and safe contact rather than maximize grip strength alone.

1X describes the hand as an API to the physical world. Better manipulation hardware allows an AI system to interact with more objects and collect richer training data.

Each grasp can produce synchronized information from cameras, joint positions, tactile sensors, force estimates and robot actions. Successful attempts show the model which actions worked. Failures show where the grasp, trajectory or force selection broke down.

This data can support end-to-end vision-action models that learn from repeated physical interaction rather than relying only on manually programmed motion.

The hardware does not remove the software problem. A dexterous hand still needs perception, task planning, collision avoidance, recovery behavior and enough training data to generalize beyond the demonstrations used during development.

1X positions NEO as a humanoid platform for residential environments. The target tasks include organizing objects, carrying groceries, cleaning, folding laundry and loading household appliances.

These tasks should be described as intended capabilities or demonstrated research tasks unless 1X has confirmed a specific commercial autonomy level.

A short edited video does not reveal intervention rate, task success rate, reset frequency or the number of attempts required. A clip should not be treated as proof that NEO can already perform unrestricted household work without supervision.

Earlier research involving NEO and NVIDIA robotics models showed the platform performing household manipulation tasks such as dishwasher loading. Coverage should remain precise about whether a specific demonstration was autonomous, teleoperated, scripted or assisted.

A human hand is commonly described as having roughly 27 degrees of freedom depending on how the wrist and smaller joint movements are counted.

NEO's 25-degree-of-freedom design approaches that mechanical range. Matching the number of joints does not automatically reproduce human dexterity.

Human manipulation also depends on dense biological sensing, soft tissue, adaptive grip strategies and years of learned coordination. A robotic hand must reproduce part of that performance through motors, tendons, sensors and control software.

Useful comparisons come from measured task performance: grasp success, slip recovery, positional accuracy, object damage, maintenance and performance on unfamiliar objects.

The next useful demonstrations should show long uncut manipulation sequences with failed grasps and recovery behavior left visible.

Relevant measurements include successful cycles, maintenance requirements, tendon replacement intervals, tactile sensor reliability and performance after extended household use.

The autonomy boundary also needs to remain clear. Viewers should know when a task is controlled by a learned policy, teleoperated by a person, executed through a scripted sequence or completed with human assistance.

If the hand maintains its precision and compliance over millions of cycles, it could remove an important hardware limitation for NEO. The remaining challenge would move toward data quality, model reliability and safe decision-making in homes that never remain perfectly organized.

1X documents 25 total degrees of freedom, low-ratio tendon transmissions, force-controlled backdrivable joints, tactile sensing, up to 45 N distal flexion force and 17.75 Nm wrist torque. Those figures come from 1X's engineering release.

The same release reports component and finger-assembly testing across millions of cycles and wrist testing beyond two million cycles under high load. Public independent tests do not yet establish long-duration household reliability, maintenance intervals or a standardized comparison with other dexterous hands.