# Shadow Robot Company

> Source: https://aiwiki.ai/wiki/shadow_robot
> Updated: 2026-07-06
> Categories: Robot Hardware, Robotics, Robotics Companies
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

| The Shadow Robot Company | |
| --- | --- |
| Type | Private limited company (UK, no. 03308007) |
| Founded | Shadow Robot Project began 1987; incorporated 27 January 1997 |
| Headquarters | London, United Kingdom |
| Key people | Richard Greenhill (founder); Rich Walker (managing director) |
| Products | Shadow Dexterous Hand, DEX-EE and DEX-EE Chiral, teleoperation systems, tactile sensors, Air Muscles |
| Employees | ~20 to 40 (2025) |
| Website | shadowrobot.com |

The **Shadow Robot Company** is a British [robotics](/wiki/robotics) firm based in London that designs and builds anthropomorphic [dexterous hands](/wiki/dexterous_hand) and teleoperation systems for research. Its flagship product, the **Shadow Dexterous Hand**, is a roughly 24-joint human-scale hand that has served as a standard laboratory platform for [robot manipulation](/wiki/robot_manipulation) for two decades, and it was the hardware behind [OpenAI](/wiki/openai)'s [Dactyl](/wiki/dactyl) in-hand manipulation experiments. Grown out of an amateur humanoid-building group founded in 1987 and incorporated as a company in 1997, Shadow is frequently described as one of Britain's longest-running robotics companies. [1][5][15]

## Origins: from an attic humanoid project to a hand company

The company's roots lie not in a business plan but in a decades-long amateur robotics project. In November 1987 Richard Greenhill, a London photojournalist with no formal engineering training, gathered about a dozen like-minded people to build a life-size humanoid robot that could do useful work such as carrying luggage. The group, which called itself the Shadow Group or Shadow Robot Project, met on Wednesday evenings, often over a pot of spaghetti cooked by Greenhill's wife Sally, and built machines from scavenged parts including components from old printers. [1][2]

Over the following years the collective produced a series of experimental machines, most famously a compressed-air-driven biped nicknamed the Shadow Walker, which used McKibben-style pneumatic "air muscles" to approximate human musculature. [2] The project became formal in 1997, when a customer asked the group to build a robot leg. To take the order they had to be a registered company, so The Shadow Robot Company Limited was incorporated on 27 January 1997. [1][15] The UK government supported the young company with an early Smart award in 1998. [5]

A strategic pivot soon followed. In 1997 Honda revealed its P2 humanoid, the product of a large, secret, well-funded program, and Greenhill concluded that a small London workshop could not compete on a full humanoid. He decided instead to build a humanoid hand, modelled on the dimensions of his own hand. [1] That decision set the company's course: rather than a walking robot, Shadow would specialise in the single hardest part of a humanoid, the hand, and in the broader problem of dexterous robot manipulation. The first fully anthropomorphic Shadow hand of the late 1990s was celebrated for a deceptively simple feat, picking up a thin-walled pint glass, and the Shadow Dexterous Hand was released commercially in 2005. [5][15]

## The Shadow Dexterous Hand

The **Shadow Dexterous Hand** is an anthropomorphic robot hand and forearm sized to match an adult human. It is one of the most detailed commercial approximations of the human hand ever sold, and its combination of high joint count, tendon drive, and dense sensing made it the default research hand for laboratories studying dexterous manipulation. [5][6]

The hand has 24 joints. Twenty of these are independently actuated [degrees of freedom](/wiki/degrees_of_freedom), and the remaining four are coupled (underactuated) movements: in each of the four fingers the two most distal joints are tendon-linked so they bend together. The thumb is fully actuated with 5 degrees of freedom across 5 joints, each finger has 3 degrees of freedom across 4 joints, and the wrist adds 2 degrees of freedom. This layout gives the hand a workspace close to that of a human hand, and in a few directions it slightly exceeds human range. [5][6]

Two actuation generations exist. The original hand was pneumatic, driven by antagonistic pairs of McKibben air muscles housed in the forearm, an approach that gave soft, compliant, muscle-like force. [3][5] Around 2010 Shadow introduced the electric "E" hand, which replaced the air muscles with 20 Maxon DC motors in the forearm that pull steel-and-polymer tendons routed into the fingers. [6] The electric hand became the version most widely used in machine-learning research because it is [tendon-driven](/wiki/tendon_driven) yet self-contained, needing no external air supply. Shadow has continued to sell both pneumatic and electric variants, plus a smaller "Lite" configuration. [6]

### Sensing and tactile feedback

A defining feature of the Shadow hand is how much it measures about itself and its contacts. A fully instrumented hand carries on the order of 120 to 129 sensors, and their data is streamed to the user at rates from 100 Hz up to 1 kHz over a high-bandwidth EtherCAT interface. [7][14][21] The onboard sensing includes absolute joint-position sensors (Hall-effect) at every joint, tendon or actuator force sensing, an inertial measurement unit, and motor current, voltage, and temperature telemetry. [5][7]

For [tactile sensing](/wiki/tactile_sensing) at the fingertips, Shadow fits its own Pressure Sensor Tactiles (PSTs) as standard, single-region pressure sensors sampled by an 11-bit converter. [8] Buyers can upgrade the fingertips to richer multimodal sensors, most notably the BioTac from SynTouch, which mimics a human fingertip and reports pressure, vibration, and temperature. [5][20] This combination of high joint count and dense touch sensing is why the hand became a research benchmark rather than an industrial tool.

### Shadow Dexterous Hand specifications

| Parameter | Shadow Dexterous Hand (E series) |
| --- | --- |
| Type | Anthropomorphic, human-scale hand and forearm |
| Joints | 24 total (20 independently actuated, 4 coupled/underactuated) |
| Thumb | 5 degrees of freedom, 5 joints |
| Each finger | 3 degrees of freedom, 4 joints |
| Wrist | 2 degrees of freedom |
| Actuation (electric) | 20 Maxon DC motors in the forearm, tendon-driven |
| Actuation (pneumatic) | Antagonistic pairs of McKibben air muscles |
| Sensors | ~120 to 129 (joint position, tendon force, IMU, motor telemetry, tactile) |
| Fingertip tactile | Pressure Sensor Tactiles (PST) standard; BioTac and other options |
| Data / control rate | 100 Hz to 1 kHz over EtherCAT |
| First commercial release | 2005 (pneumatic); electric "E" hand ~2010 |

## Role in AI research: OpenAI's Dactyl and the Rubik's cube

Shadow's place in the history of AI comes chiefly from OpenAI's Dactyl project, which used a Shadow Dexterous "E" hand as its physical platform. In the 2018 "Learning Dexterity" work, OpenAI trained a control policy entirely in simulation to reorient a block held in the hand, then transferred it to the real Shadow hand. [9] The key idea was [sim-to-real](/wiki/sim_to_real) transfer with domain randomization: by training across thousands of randomized simulated variations of physics, appearance, and sensor noise, the learned policy became robust enough to work on real hardware it had never touched during training. [9][11]

In October 2019 OpenAI extended the work to a far harder demonstration: manipulating a Rubik's cube one-handed on the same Shadow hand. [10] An important accuracy point is division of labour. A classical algorithm decided which cube moves to make, while the neural network, trained with [reinforcement learning](/wiki/reinforcement_learning), controlled the physical dexterity of turning faces and reorienting the cube in the fingers. To make simulation-trained policies survive contact with reality, OpenAI introduced Automatic Domain Randomization (ADR), which progressively increased the difficulty and variety of the simulated world as the policy improved. [10][11] The physical setup added a PhaseSpace motion-capture system and RGB cameras for pose estimation, and OpenAI modified the hand by moving tracking LEDs and cables inside the fingers and adding rubber to the fingertips. [10] Reported success rates were roughly 60 percent on fairly scrambled cubes and about 20 percent on a maximally scrambled cube, and the policy kept working under perturbations it was never trained on, such as a rubber glove, tied-together fingers, or being prodded with a stuffed giraffe. [10][11]

Dactyl and the Rubik's cube became a landmark in learned robot manipulation and helped popularize simulation-first [robot learning](/wiki/robot_learning). They also exposed a hardware limitation that would shape Shadow's next product: a hand built to imitate human anatomy was fragile under the millions of collisions of trial-and-error learning, and OpenAI's team spent significant effort keeping the delicate 20-motor hand running. [10][17]

## DEX-EE: a hand built for reinforcement learning

The lesson that research-grade [deep learning](/wiki/deep_learning) needs indestructible hardware led directly to Shadow's most important recent product. Google DeepMind's robotics team approached Shadow in 2017 wanting a hand that could learn on real-world tasks yet survive the punishment of long, unattended reinforcement learning runs. [16] The result, unveiled with [Google DeepMind](/wiki/google_deepmind) and announced by Shadow on 30 January 2025, is the **DEX-EE**, a three-fingered research hand engineered for robustness first. [15]

DEX-EE trades the full 24-joint anatomy of the classic hand for a simpler, tougher design: three fingers, 12 degrees of freedom, a weight of about 4.1 kg, a height of about 350 mm, and roughly 15 Maxon DCX16 DC motors, with the whole hand around 50 percent larger than a human hand. [14][15] Each finger carries hundreds of channels of tactile data, using optical fingertip sensors with hundreds of taxels each plus multi-taxel sensors on the middle and proximal segments, alongside position, force, and inertial sensing throughout. [14][16] Crucially for AI work, the tendons are designed to absorb repeated hard impacts even when several are loaded at once, the fingers and parts are quickly swappable, and the system includes fail-safes and a graceful shutdown routine. Shadow says the hand went through more than 1,000 hours of testing, including deliberate impact and shock testing with pistons and tools, to reach that durability. [15][22]

DEX-EE is the hand behind DeepMind's 2024 dexterity research. In September 2024 DeepMind published DemoStart, a method that uses an auto-curriculum and a handful of demonstrations to train multi-fingered manipulation in simulation and transfer it to the real DEX-EE. [12][13] DemoStart reported over 98 percent success in simulation and about 97 percent on real-world cube reorientation and lifting, and 64 percent on a plug-and-socket insertion task requiring fine finger coordination, while needing roughly 100 times fewer simulated demonstrations than comparable real-world learning. [12][13] The DEX-EE is now sold commercially as a research platform, and Shadow has added the **DEX-EE Chiral**, a variant whose third finger is offset "down and to the side" to imitate a human thumb, which makes teleoperation and [imitation learning](/wiki/imitation_learning) more natural. The Chiral is offered in left, right, and bi-manual pair configurations. [14]

| Parameter | DEX-EE | DEX-EE Chiral |
| --- | --- | --- |
| Fingers | 3 (symmetric) | 3 (human-like thumb offset) |
| Degrees of freedom | 12 | 12 |
| Weight | ~4.1 kg | ~4.1 kg |
| Height | ~350 mm | ~350 mm |
| Actuators | ~15 Maxon DCX16 DC motors | ~15 Maxon DCX16 DC motors |
| Tactile sensing | Optical fingertips (hundreds of taxels each) + multi-taxel middle/proximal sensors | Same, plus configurations tuned for imitation learning |
| Robustness | 1,000+ hours of testing; impact/shock tested; fail-safes | Same platform |
| Availability | Commercial research platform | Left, right, bi-manual pair |

## Teleoperation and haptics

Alongside autonomous manipulation, Shadow builds teleoperation systems that let a human operator control a robot hand remotely and feel what it touches. The teleoperation product pairs a Shadow Dexterous Hand (marketed with 20 degrees of freedom, 24 movements, and about 120 sensors) with either Shadow's own in-house glove or a HaptX glove fitted with Shadow tactile fingertips, so the operator receives pressure, temperature, and vibration feedback. Target uses include nuclear decommissioning, pharmaceutical handling, maintenance, and research in hazardous environments. [17][21]

The most visible teleoperation effort was the Tactile Telerobot, built by the Converge Robotics Group, an international collaboration of Shadow Robot, haptics company HaptX, and Tangible Research, using SynTouch BioTac sensors and Universal Robots arms. In a widely reported 2019 demonstration an operator in California typed on a keyboard in London and felt each key press, and the system went on to reach the finals of the 10 million dollar ANA Avatar XPRIZE. [20] At ICRA 2026 Shadow demonstrated teleoperation of both its flagship Shadow Hand and the DEX-EE Chiral using MANUS Metagloves Pro Haptic gloves. [21]

## ARIA-funded research and users

In 2025 Shadow became a lead recipient in the UK Advanced Research and Invention Agency (ARIA)'s Robot Dexterity program. On 29 May 2025 ARIA announced roughly 11 million pounds of funding for two Shadow-led projects: OGRES (Optimised General Robot End-effector System), which aims to build software tools that take a hand from simulated design to a working physical hand, with partners including the Universities of Cambridge, Imperial, Bristol, and Bath, EPFL, MorphoAI, Basis, and VSim; and UPWARD (UnPrecedented actuators: Paving the Way for Advanced Robotic Dexterity), which tackles how to power every joint of a hand without bulking it up, with partners including Northwestern University, Texas A&M, and the UK Manufacturing Technology Centre. [18] The awards sit inside ARIA's broader Robot Dexterity program, which is investing on the order of 57 million pounds across dozens of teams. [19]

Documented users of Shadow hands over the years include NASA, Carnegie Mellon University, Bielefeld University, and the European HANDLE research project, in addition to OpenAI and Google DeepMind. [5] Shadow hands have also been applied in [physical AI](/wiki/physical_ai) and hazardous-environment work such as pharmaceutical lab automation and nuclear decommissioning. [17][21]

## Current status and ownership

As of mid-2026 The Shadow Robot Company Limited remains an independent, privately held UK company headquartered in London, with an active status on the UK companies register (no. 03308007) and a small team of roughly 20 to 40 people. [15][6] Its founder Richard Greenhill is still associated with the company, and day-to-day leadership has long rested with managing director Rich Walker, an engineer who joined the project as a teenager before it was incorporated. [1][15]

No acquisition, buyout, or insolvency has been reported through 2025 and 2026. Public databases record the company as sustained by a mix of product sales, industrial and research collaborations (most prominently with Google DeepMind), and public research grants rather than large venture rounds; its most significant recent funding is the 2025 ARIA award. [18][24] The company was visibly active in 2026, exhibiting and demonstrating both the Shadow Hand and DEX-EE Chiral at ICRA 2026. [21] Any single market-size or valuation figure for Shadow should be treated with caution, because the company is small, privately held, and does not publish detailed financials beyond its UK statutory filings.

## ELI5

Imagine a robot hand that has the same fingers and joints as your hand, with tiny motors and strings inside that work like muscles and tendons, and touch sensors on the fingertips so it can feel what it is holding. That is the Shadow Dexterous Hand, made by a small company in London that started in 1987 as a hobby group trying to build a whole robot person. They discovered the hand is the hardest part, so they focused on it. When companies like OpenAI and Google DeepMind wanted to teach robots to move their fingers by practicing, they used Shadow's hands, including a tougher three-fingered one called DEX-EE that can be bashed around thousands of times while it learns.

## See also

- [Dexterous hand](/wiki/dexterous_hand)
- [Humanoid robot hands](/wiki/humanoid_robot_hands)
- [Tactile sensing](/wiki/tactile_sensing)
- [Tendon-driven](/wiki/tendon_driven)
- [Robot manipulation](/wiki/robot_manipulation)
- [Robot learning](/wiki/robot_learning)
- [Dactyl](/wiki/dactyl)
- [Allegro Hand](/wiki/allegro_hand)
- [Google DeepMind](/wiki/google_deepmind)

## References

1. Shadow Robot Company, "The story of our Founder, Richard Greenhill." https://www.shadowrobot.com/blog/the-story-of-our-founder-richard-greenhill/
2. Evan Ackerman, "Shadow Walker Was a DIY Biped Humanoid Robot," IEEE Spectrum. https://spectrum.ieee.org/shadow-walker-biped-humanoid-robot
3. Researchers, "McKibben pneumatic actuators (manufactured by Shadow)," ResearchGate figure and review literature on McKibben air muscles. https://www.researchgate.net/figure/McKibben-pneumatic-actuators-relaxed-top-and-inflated-bottom-Manufactured-by-Shadow_fig1_224441600
4. techSPARK, "25 years strong for a robotics company with a human approach," January 2022. https://techspark.co/blog/2022/01/27/25-years-strong-for-a-robotics-company-with-a-human-approach/
5. Wikipedia, "Shadow Hand." https://en.wikipedia.org/wiki/Shadow_Hand
6. Shadow Robot Company, "Dexterous Hand Series." https://shadowrobot.com/dexterous-hand-series/
7. Shadow Robot Company, "Shadow Dexterous Hand E Technical Specification" (PDF). https://www.shadowrobot.com/wp-content/uploads/2022/03/shadow_dexterous_hand_e_technical_specification.pdf
8. Shadow Robot Company, Dexterous Hand documentation and FAQ (Pressure Sensor Tactiles). https://shadow-robot-company-dexterous-hand.readthedocs-hosted.com/en/2.1.5/user_guide/4_FAQ.html
9. OpenAI, "Learning Dexterity," 2018. https://openai.com/index/learning-dexterity/
10. OpenAI, "Solving Rubik's Cube with a Robot Hand," October 2019; arXiv:1910.07113. https://arxiv.org/pdf/1910.07113
11. Evan Ackerman, "OpenAI Demonstrates Sim2Real by Solving a Rubik's Cube One-Handed," IEEE Spectrum. https://spectrum.ieee.org/openai-demonstrates-sim2real-by-with-onehanded-rubiks-cube-solving
12. Google DeepMind, "Our latest advances in robot dexterity," September 2024. https://deepmind.google/blog/advances-in-robot-dexterity/
13. Bauza et al., "DemoStart: Demonstration-led auto-curriculum applied to sim-to-real with multi-fingered robots," arXiv:2409.06613, 2024. https://arxiv.org/abs/2409.06613
14. Shadow Robot Company, "DEX-EE Series." https://shadowrobot.com/dex-ee_series/
15. Mike Oitzman, "Shadow Robot DEX-EE hand takes manipulation to next level," The Robot Report, 30 January 2025. https://www.therobotreport.com/shadow-robot-dex-ee-hand-takes-manipulation-to-next-level/
16. Wevolver, "Gripping the Future: Shadow Robot's DEX-EE and the Next Generation of Dexterous Manipulation." https://www.wevolver.com/article/gripping-the-future-shadow-robots-dex-ee-and-the-next-generation-of-dexterous-manipulation
17. techjournal.uk, "Shadow Robot Builds Hand That Learns Through Touch and Impact." https://www.techjournal.uk/p/shadow-robot-builds-hand-that-learns
18. Shadow Robot Company, "ARIA announces funded projects with Shadow Robot Company," 29 May 2025. https://shadowrobot.com/aria-announces-funded-projects-with-shadow-robot-company/
19. Richard Speed, "UK's DARPA clone invests 23.3M pounds in touchy-feely robots," The Register, 30 May 2025. https://www.theregister.com/2025/05/30/uks-darpa-clone-invests-233m-in-touchy-feely-robots/
20. HaptX, "Tactile Telerobot reaches finals of 10 Million dollar ANA Avatar XPRIZE." https://haptx.com/tactile-telerobot-reaches-finals-of-10-million-ana-avatar-xprize/
21. Shadow Robot Company, "Teleoperation" and "ICRA 2026 Recap." https://shadowrobot.com/teleoperation/ and https://shadowrobot.com/icra-2026-recap/
22. RoboticsTomorrow / maxon group, "The most sensitive and durable robot hand yet created," January 2025. https://www.roboticstomorrow.com/story/2025/01/the-most-sensitive-and-durable-robot-hand-yet-created/24025/
23. Luke Dormehl, "The holy grail of robotics: Inside the quest to build a mechanical human hand," Digital Trends. https://www.digitaltrends.com/cool-tech/shadow-robot-company-hand/
24. THE SHADOW ROBOT COMPANY LIMITED, overview and filing history, UK Companies House (no. 03308007). https://find-and-update.company-information.service.gov.uk/company/03308007