Dexterous Robot Hand: Stanford, Shadow and ABB GoFa

A dexterous robot hand has several independently driven fingers, so it can turn, pinch and regrip an object, where a two-jaw gripper can only clamp it. Engineers have been building them for more than 40 years, and in October the market researcher TrendForce called the hand the hardest part of a humanoid robot to get right. This post goes from the Stanford/JPL and Utah/MIT hands of the 1980s through the Shadow Hand and NASA's Robonaut, then looks at this autumn's new hands and at ABB's tests of a prosthetic hand on its GoFa cobot.

A close-up of a hand of Robonaut 2, NASA's dexterous humanoid robot, on the International Space Station in February 2012. Photo: NASA (public domain)

How robot hands went from grippers to the Shadow Hand

The first robot hands of the 1960s and 1970s were grippers built for industrial jobs, with little dexterity. The Stanford/JPL hand of the early 1980s, usually called the first dexterous robot hand, was built to study fine manipulation. It had three independently controlled fingers with nine degrees of freedom in total and force feedback, and one setup added fingertip load cells for touch sensing. Kenneth Salisbury, who finished his Stanford PhD in mechanical engineering in 1982, worked on its development. His later haptics work led to SensAble Technology, maker of the PHANToM haptic interface, and he also worked on the da Vinci surgical robot.

A Unimate PUMA robot with a simple gripper beside a milling machine in a computer integrated manufacturing cell, 1986. Photo: UL Digital Library via Wikimedia Commons (no known restrictions)

The Utah/MIT Dextrous Hand followed soon after and pushed human-like articulation further. A project video catalogued by the Computer History Museum shows a single finger, a full hand, the hand mounted on a PUMA 500 arm and a later version labelled the Gold Hand. The hand had no tactile sensing. In the 1990s researchers turned to underactuated hands, which use fewer actuators than joints to keep control simple. NASA and DARPA meanwhile built Robonaut at Johnson Space Center, a humanoid whose hands were sized for the strength spacewalk tasks need, and its successor R2 reached the International Space Station on one of the last shuttle flights in 2011.

NASA's Robonaut anchored to a test panel at Johnson Space Center with both of its dexterous hands free, 2004. Photo: NASA (public domain)

In London, the Shadow Robot Company grew out of a group of hobbyists meeting in its founder's attic. Its first wooden hand appeared in 1997, a university asked to buy it, and the business formed around that order. An air muscle version became the flagship in 2004, and a motor-driven Shadow Dexterous Hand replaced it in 2010. The hand has 24 joints and 20 degrees of freedom, more than a human hand, driven either by 20 DC motors or by 20 antagonistic pairs of air muscles in the forearm, with Hall effect sensors in every joint. NASA, Bielefeld University and Carnegie Mellon University are among its users.

A Shadow Dexterous Hand holding a light bulb, 2007. Photo: Richard Greenhill and Hugo Elias, Shadow Robot Company / Wikimedia Commons (CC BY-SA 3.0)

Dexterous hands in 2026: humanoids, new hardware and ABB

TrendForce's report of 22 October quotes Elon Musk calling human-like robot hands "shockingly difficult". It says the hands of Tesla's Optimus still depend heavily on manual assembly, that their tactile sensors are still being made more reliable, and that Tesla plans a modular, swappable sensing glove for its 2027 model. TrendForce expects hands to make up 20 to 30 percent of a humanoid's hardware cost over the long term. It projects that humanoid sales will pass 2 million units by 2030, which could push the hand market to about US$6 billion. China had nearly 510 companies working on dexterous hands in the first half of 2026.

Robonaut 2 shakes hands with NASA astronaut Dan Burbank on the International Space Station in 2012, the first human and robot handshake in space. Photo: NASA (public domain)

At IROS 2026 in Pittsburgh, held from 27 September to 1 October, Inspire Robots unveiled the RH5MK1, a fully direct-drive hand at human scale with 22 active degrees of freedom. Its fingers open and close five times a second, and 6D fingertip tactile sensors resolve forces down to 0.005 N. Inspire's RH56E2 is aimed at industry, with 30 N of fingertip grip force and a 15 kg payload for the whole hand. Shadow, for its part, now sells DEX-EE, a three-fingered hand it developed with Google DeepMind for machine learning research, with optical fingertip sensors that each carry hundreds of tactile sensing points.

A Robonaut hand grasps a tether hook during NASA evaluations in 2004. Photo: NASA (public domain)

In June 2026 ABB Robotics said it was pairing PSYONIC's Ability Hand, a prosthetic hand with myoelectric control, touch sensing and compliant fingers, with its GoFa cobot. The two companies want to find out whether touch and motion data from people who wear the hand can train robots for delicate, variable tasks, with GoFa supplying the accuracy and repeatability to repeat and compare small differences in grip force and finger position. Our post on prosthetic hands covers the prosthetic side. ABB's dual-arm YuMi already uses servo grippers that ABB calls its hands, with optional built-in cameras, on two seven-axis arms.

Two ABB GoFa cobots fitted with PSYONIC Ability Hands, one holding a torque wrench on a metal drive housing. Photo: ABB

Where dexterous hands go next, and what still holds them back

PSYONIC founder Aadeel Akhtar calls dexterous manipulation "a data challenge as much as a hardware challenge", and the new hardware is designed around collecting that data. Shadow built DEX-EE to run long machine learning experiments without hardware failures, and Inspire gave the RH5MK1 joints with active zero-position calibration to limit sensor drift during continuous data collection. ABB is working on the software side. In October it added an AI Robot Trainer to OmniCore EyeMotion, which turns shop floor images into AI vision models and collision-free robot paths, and ABB says the combination cuts commissioning time by up to 90 percent.

Overhead view of an ABB cobot picking mixed items from trays, from ABB's October 2026 OmniCore EyeMotion AI Robot Trainer release. Photo: ABB

ABB and PSYONIC say they will test industrial jobs where ordinary grippers struggle with variable, fragile or complex parts, in automotive, aerospace, packaging and logistics, and life sciences. ABB's release cites an International Federation of Robotics estimate that advanced gripping and digital integration can cut engineering time by up to 30 percent. ABB puts grasping and dexterity at the centre of what it calls Autonomous Versatile Robotics, its plan for robots that sense, reason and handle objects precisely in changing surroundings. Factory cells built around such hands would usually be planned and tested first in RobotStudio.

GoFa cobots with Ability Hands picking coloured blocks from a wooden box. Photo: ABB

At IROS, Inspire's strategy head Liu Wenrui named four pain points for the whole industry: reliability, testing standards, compatibility between core components, and cost, down to details such as fatigue life and load testing. On cost, TrendForce ranks the hand as the most expensive part of a humanoid after the body's actuators, with tactile sensors, coreless motors, reduction gears and lead screws among its costliest components. Investors are paying for the attempt anyway: three hand module makers in China, LinkerBot, AGILINK and GeniTech, raised RMB 5.3 billion, about US$790 million, in the first half of 2026.

The hands of Robonaut 2 during initial checkouts in the Destiny laboratory of the International Space Station, 2012. Photo: NASA (public domain)

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com