Robot hands: dexterity, sensing and evidence

End-effectors should be selected for the work they must perform. A parallel gripper can be reliable for repetitive pick-and-place, while a multi-finger hand can support more grasp shapes, in-hand motion, and tools designed for people. Adaptive or underactuated hands sit between these categories by allowing several joints to conform with fewer independently controlled actuators.

The words hand, gripper, dexterous, and anthropomorphic do not establish capability by themselves. A defensible comparison names the joints and actuators, sensing, payload and object range, control interface, cycle time, durability, task protocol, and failure behavior actually tested.

A robot hand is a coupled mechatronic and software system. Mechanical design sets reachable contact configurations; actuation and transmission determine controllability; sensors expose joint and contact state; the controller turns those signals into coordinated motion; and the policy or planner selects actions for a task.

Degrees of freedom and actuator count describe architecture, not universal dexterity. Payload, fingertip force, speed, repeatability, tactile coverage, compliance, environmental tolerance, power, mass, maintenance, software access, and task evidence all matter. Values from different test methods should not be placed in one ranked table without aligning definitions and conditions.

For procurement or research selection, record whether each specification is a manufacturer statement, a calibrated measurement, a peer-reviewed result, an independent benchmark, or an observation from a demonstration. Unknown fields should stay unknown rather than being inferred from a product image or marketing name.

A fingertip sensor can resolve local contact for insertion, slip response, or texture-related tasks. Finger and palm sensing can expose load paths and contacts that a fingertip-only layout misses. Whole-hand systems increase coverage but create routing, calibration, durability, bandwidth, and representation challenges.

The Nature Machine Intelligence full-hand tactile sensing work and the HT-Bench/HandTouch preprint are useful research examples, but they answer different questions. The first demonstrates an integrated full-hand sensing approach; the second proposes data and evaluation tracks for learned full-hand representations. Neither source proves that one sensor or representation is best for every hand and task.

Google DeepMind’s Gemini Robotics 2 announcement reports experiments across whole-body Apollo hardware, a multi-finger Sharpa hand, and a Franka Duo gripper setup. Those are official developer-reported evaluations, not an independent cross-hand benchmark. They show why embodiment and end-effector must remain attached to every task result.

TactiDex, HRDexDB, and related preprints investigate tactile skill transfer or reusable hand data. Their datasets, robots, sensors, object sets, and protocols differ, so reported results should remain source-bounded rather than being converted into a universal hand ranking.

ADEPT compares two different arm-hand workbenches but does not train one cross-hand checkpoint. Its KUKA-Allegro student is vision-only; the five-fingertip tactile pathway and 3/10-versus-8/10 matched result belong only to the Flexiv-Sharpa square-and-round insertion condition.

A useful tactile-hand evaluation starts with a task that genuinely depends on contact: occluded grasping, slip, insertion, reorientation, deformable objects, handover, or disturbance recovery. It then aligns the hand, objects, sensors, controller, trial count, success definition, and no-touch baseline.