Robot hands are shipping and getting better. What is known about hands, touch and dexterity, who argues what, and what comes next, with a source for every fact.
Robot hands moved a long way in the last twelve months. Boston Dynamics showed a four-finger Atlas hand built for tools and for mass manufacturing (Boston Dynamics, 1 Oct 2026); 1X published a 25-degree-of-freedom hand with tactile skin, wrist joints tested beyond 2 million cycles and capacity for 10,000 hands a year (1X, 9 Jul 2026); and a market of hands you can buy now exists outside the big humanoid makers (CNX Software, 2 Jun 2026). Learning moved too: Physical Intelligence won Benjie Holson's manipulation "Olympics" with cameras and two-finger grippers within about three months (General Robots, 22 Jan 2026), and a 2026 policy that gives touch its own fast loop almost doubled success on twelve real tasks (arXiv 2606.17055). The next step is the one buyers ask about: no hand maker has yet published how long a hand lasts in a working shift, and Tesla's reported production problems sit mostly in the hand (The Next Web, 28 Sep 2026).
How many fingers does useful work need?
Boston Dynamics moved Atlas from 3 fingers and 7 degrees of freedom to 4 fingers and 13, and left out the little finger because "an extra finger means three more actuators, with the corresponding extra cost, volume, and probability of something breaking" (Boston Dynamics, 1 Oct 2026).
Three-finger hands are now products: Shadow Robot's DEX-EE (3 fingers, 12 degrees of freedom, built with Google DeepMind) is on sale (Shadow Robot), and Tangent Robotics raised $4.5 million on 30 September 2026 for a three-finger hand aimed at assembly work (Humanoids Daily, 30 Sep 2026).
Five-finger humanoid hands are the other bet: Tesla's Optimus hand has 22 degrees of freedom with its actuators in the forearm (The Decoder, 28 Nov 2024), and 1X's NEO hand has 22 fully actuated joints in fingers and palm (1X, 9 Jul 2026).
Tendons from the forearm, or motors in the finger?
The tendon camp: Optimus runs tendons through the wrist from forearm motors (Drive Tesla Canada, 17 Apr 2026); 1X drives its tendons at low gear ratios of about 5:1 to 15:1 so every joint is backdrivable and force-controlled (1X, 9 Jul 2026); mimic's M1 hand (launched 16 July 2026) is tendon-driven from the forearm and holds more than 25 kg (SiliconANGLE, 16 Jul 2026).
The in-finger camp: the new Atlas hand uses one actuator type, driven at the joints, "completely encapsulated, with no fragile cables crossing joints" (Boston Dynamics, 1 Oct 2026).
Tendons can last: the open-source Aero Hand's cable drive survived more than 400,000 full closing cycles on the bench, with parts costing $314 (arXiv 2608.28578).
How much touch does a hand need?
Touch is in products: Figure 03's fingertips detect about 3 grams, with a camera in each palm (Figure, 9 Oct 2025); Atlas has pressure sensors on fingertips and palm (Boston Dynamics, 1 Oct 2026); Contactile reports more than 100 customers in more than 12 countries for sensors that compute grip force on the sensor (IT Brief, 9 Jul 2026).
Touch gives measured gains when it is used the right way. T-Rex, with a separate fast touch loop, reached 65% average success on 12 real tasks against 35% for the best vision-based model; simply adding touch to a standard model made it worse (6% against 17%) (arXiv 2606.17055). TacVLA reports a 2.1x gain when the camera is blocked (arXiv 2603.12665).
Some of the gain survives without a sensor at run time: TacImag trains with touch and deploys with cameras only, because of "the fragility, calibration requirements, and maintenance burden of tactile hardware" (arXiv 2607.01684).
Can hands be built, and repaired, at scale and at a sane cost?
Production is starting: SharpaWave hands began shipping in October 2025 in rolling mass production (Interesting Engineering, 19 Dec 2025); 1X makes tendons, motors, skin and touch sensors in-house, with capacity for 10,000 hands a year (1X, 9 Jul 2026); PSYONIC's Ability Hand is worn by more than 300 prosthetic patients and used by more than 90 robotics companies, and its 2025 revenue grew to $5.35 million (Runtime Wire, 22 Jun 2026).
Makers now publish durability numbers: SharpaWave claims 1 million grip cycles (Interesting Engineering, 19 Dec 2025); Shadow Robot reports more than 1,000 hours of testing of DEX-EE, including impact tests (maxon). These are company claims; no field failure rate has been published.
Tesla says the hand is its hardest part: on the Q3 2025 call Musk described the hand and forearm as harder than the rest of the robot (live notes by Shacknews, 22 Oct 2025). The Information reports more than 100 hand-assembled screws and parts per hand and forearm, and output rising from a few dozen robots a week in the second quarter of 2026 to several hundred a week in August (as relayed by The Next Web, 28 Sep 2026; not confirmed by Tesla).
Does dexterity come from data, from simulation, or from design?
Human video scales: NVIDIA's EgoScale trained on 20,854 hours of egocentric human video, found a log-linear scaling law and raised success by 54% on a 22-degree-of-freedom hand (arXiv 2602.16710).
Wearables help: DexUMI puts an exoskeleton on the human hand so the person feels the contact, and reached 86% average success on two robot hands (arXiv 2505.21864); mimic's U1 glove records data without a robot (eWeek, 22 Jul 2026).
Simulation is reaching real hands: a January 2026 paper moved grip-force control and in-hand reorientation from simulation to a real five-finger hand with no fine-tuning (arXiv 2601.02778); Boston Dynamics calls its first sim-to-real results on the new hand "promising" and "initial" (Boston Dynamics, 1 Oct 2026).
When do simple grippers or suction beat a hand?
Amazon's Vulcan handles about 75% of the item types in its fulfillment centers with paddles, a blade and a suction cup, all with force sensing and no hand; the average stowing robot is now slightly faster than the average human stower, after more than 500,000 stows in Spokane and Germany (IEEE Spectrum, 7 May 2025).
Industry still runs on parallel-jaw grippers and suction; Brooks notes Schunk alone sells more than 1,000 kinds of parallel-jaw gripper (rodneybrooks.com, 26 Sep 2025).
The gripper results above (Holson's Olympics, Gemini Robotics 2) show how far two fingers and good learning go.