What is a dexterous robot hand?

Why it is so hard

A gripper has one degree of freedom and works reliably on objects it was designed around. A dexterous hand aims at everything a human hand does, which means many joints, in a small volume, that must also survive being hit.

Everything competes for the same space: actuation, transmission, wiring, cooling, and tactile sensing. Add a finger and you add a motor, a cable route, a sensor run and a failure point. The hand is also the part most exposed to impact — the part of the robot that touches the world by design.

The tendon argument

The central hardware dispute. Tendon-driven hands put the motors in the forearm and pull cables to the fingers, the way your own forearm works. Mass leaves the extremity, which makes the fingers fast and light. The costs are friction, cable stretch, wear and control that is harder to make precise.

Direct-drive puts a motor at each joint. Heavier fingers, but no cables to stretch and a much cleaner path from commanded torque to force at the fingertip.

Figure's Brett Adcock dismissed tendons as a complete local maximum, igniting a public fight with Foundation over friction, force and the future of humanoid hands. It is a genuine disagreement about which problem is easier to engineer away: cable friction, or the mass of putting motors in the fingers.

AGILINK has bet on direct drive at the high end, debuting the OmniHand 3 Ultra-M with a 20-DoF architecture and vision-based tactile sensing.

Touch is the actual bottleneck

Grasping without touch is grasping blind. You know a glass is slipping before you have dropped it, and you know it from your fingertips rather than your eyes. Robots mostly do not.

That is why vision-based tactile sensing — a camera inside the fingertip watching a deformable pad — keeps appearing: it borrows a mature, cheap component to get a hard signal. It is not solved. Durable, affordable touch across a whole hand remains an open problem, and it gates the manipulation everything else depends on.

An adjacent route is to borrow from prosthetics, where multi-finger hands have shipped commercially for years: PSYONIC's Ability Hand became a native asset in NVIDIA Isaac Lab, with a real-to-real transfer pipeline using human-driven data.

The money arriving

Hands have become a business rather than a component. AGIBOT spun out AGILINK, which passed a one-billion-dollar valuation within months to mass-produce dexterous hands. 1X has revealed 25-DoF hands for NEO, framing them as the final boss of home robotics.

What the money buys is iteration. Figure's first look at its seventh-generation hand is pitched at parity with a human one, adding thumb rotation and finger abduction. Seventh is the number worth noticing: nobody reaches a seventh revision of a subsystem that was already working.

Reading a hand spec

DoF is not fingers, and not always actuators. A hand can quote 20 degrees of freedom with far fewer motors, because joints are coupled — one motor curling a whole finger. Ask how many are independently driven.

Ask what it can feel. Fingertip force, slip detection, distributed skin, or nothing.

Ask what it survives. A hand that manages delicate assembly and breaks when the arm swings into a table frame is a laboratory result.

Ask whether it was driven. Almost every impressive hand video is teleoperated, and that is fine — but it is a claim about the mechanism, not about autonomy.