Shadow Dexterous Hand

High-DoF tendon-driven dexterous robot hand for research

Shadow Dexterous Hand

Home > Hardware > Hands > Shadow Hand

Overview

ItemSpecSource
ManufacturerShadow Robot Company (London, UK)1
Founded19871
Commercialized20052
DoF20 degrees of freedom, 24 movements3
Motors203
Sensors129+ (up to 1kHz sampling)3
Drive MethodTendon-driven3
CommunicationEtherCAT (100Mbps)3
SoftwareFull ROS integration3
PriceApproximately 110,000 euros (including installation, training, support, as of 2023)4

Key Significance

Shadow Dexterous Hand is one of the most widely adopted standard platforms in the dexterous manipulation research field. 2

  • Highest Level DoF: One of the few robot hands on the market offering 24 movements and 20 DoF 3
  • Human-Hand-Like Structure: High flexibility with each finger able to move laterally independently 3
  • Long Track Record: Since commercialization in 2005, adopted by major research institutions including NASA, OpenAI, Google DeepMind, and Carnegie Mellon 2
  • Major Platform for AI/ML Research: Widely used in reinforcement learning-based dexterous manipulation research 56

Design Philosophy

Biomimetic Design

The core philosophy of Shadow Hand is faithfully reproducing the anatomical structure of the human hand. 3

Tendon-Driven System

  • Transmits force through tendons like the human musculoskeletal system
  • Implements natural and flexible movement
  • Enables independent yet coordinated control of each joint
  • Continuous force transmission from wrist to fingertips

Sensor Architecture

  • 129+ sensors mimic human hand proprioception and tactile sense 3
  • Provides various sensory information including position, force, and temperature
  • High-speed sampling up to 1kHz enables real-time control

Finger Structure

  • 5 fingers (including thumb, 4 fingers + opposable thumb)
  • Independent adduct/abduct movement for each finger
  • Joint structure similar to human

Detailed Specs

Drive System 3

ItemSpecification
Motor Count20
Torque Loop Frequency5kHz (inside motor unit)
Control MethodPosition control (host PC), ROS-based custom control possible

Sensor System 3

Sensor TypeDescription
PST (Pressure Sensor Tactiles)Standard equipped, fingertip pressure detection, 11-bit ADC
BioTac Sensors (option)SynTouch LLC partnership, force/micro-vibration/temperature detection 7
Joint Position SensorsDetects each joint position
Force SensorsMeasures tendon tension

BioTac Sensor Details 7

  • 24 taxels (tactile elements)
  • 100Hz sampling
  • Overall fluid pressure and temperature change detection
  • Form and mechanical properties similar to human fingertips

Communication and Interface 3

  • EtherCAT Bus: 100Mbps Ethernet-based industrial communication
  • Full ROS Integration: ROS nodes and messages provided
  • Self-Contained System: All actuation and sensing built into hand and forearm

Product Lineup

Shadow Dexterous Hand E Series 3

Standard Research Model

  • 24 movements, 20 DoF
  • 5-finger structure
  • 129+ sensors
  • ROS integration
  • Price: Approximately 110,000 euros (approximately $180,000 with BioTac, as of 2023) 4

Shadow DEX-EE Series 8

For Long-Duration Reinforcement Learning Experiments (Developed in collaboration with Google DeepMind)

DEX-EE was developed over 5 years at Google DeepMind’s request, a next-generation hand optimized for reinforcement learning experiments. 8

ItemSpec
Finger Count3 (robust 3-finger design)
SizeApproximately 50% larger than human hand
Motors15 maxon DCX16 DC motors
Durability Testing1,000+ hours

Sensor System (per finger) 8

  • 5 tendon force sensors
  • 5 motor encoders
  • 4 joint angle sensors
  • 3 IMUs (Inertial Measurement Units)
  • 36 taxels (middle/proximal phalanx tactile sensors)
  • 50 FPS 640x480 stereo video (distal phalanx tactile sensor)

Design Features

  • Withstands repeated impacts and aggressive use during policy learning
  • Modular structure supports alternative finger layouts
  • Smooth torque control at all joints
  • Long-duration continuous experiments without hardware failure

Major Research Cases

1. OpenAI: Learning In-Hand Object Manipulation (2018)

Paper: “Learning Dexterous In-Hand Manipulation” 5

  • Trained with reinforcement learning in simulation, then transferred to real Shadow Hand
  • Important research in Sim-to-Real Transfer field
  • Achieved 13 consecutive rotations median in object reorientation task
  • Natural behaviors like finger gaiting, multi-finger coordination emerged without human demonstration

2. OpenAI: Solving Rubik’s Cube with Robot Hand (2019)

Paper: “Solving Rubik’s Cube with a Robot Hand” 6

  • Solving Rubik’s Cube with single robot hand - unprecedented complexity manipulation
  • Developed Automatic Domain Randomization (ADR) algorithm
  • Training scale: 64 NVIDIA V100 GPUs, 920 worker machines
  • Cumulative experience: approximately 13,000 years (similar scale to OpenAI Five)
  • Development period: May 2017 ~ October 2019 (approximately 2.5 years)

3. Google DeepMind: DEX-EE Development Collaboration 8

  • 5-year collaboration with Shadow Robot to develop DEX-EE
  • Solved hardware durability issues in reinforcement learning experiments
  • Achieved both rich sensor data and robustness

4. Other Major Research

Grasping in the Dark (ICRA 2021) 9

  • Grasping objects of various shapes/sizes/weights using only BioTac tactile sensors
  • Closed-loop grasping with classical control without prior knowledge

Google Brain: Deep Dynamics Models 10

  • Multi-object manipulation learning with only 4 hours of real robot data
  • Task planning based on DDM (Deep Dynamics Models)

Limitations and Considerations

LimitationDescription
High PriceOver 110,000 euros (research budget acquisition required) 4
MaintenanceRegular management of tendon-based system required
Learning CurveInitial setup and operation learning needed for complex system
Experiment EnvironmentPre-DEX-EE models vulnerable to long-duration RL experiments

Shadow Robot Company Introduction

ItemDetailsSource
Founded1987 (Richard Greenhill, started in London attic)1
Official Registration1997 (triggered by robot leg component commission)2
HeadquartersLondon, UK1
BranchesBristol, Madrid1
NotableOne of the UK’s longest-running robot companies1

Major Customers: NASA, ESA, OpenAI, Google DeepMind, Carnegie Mellon, UCL, University of Bielefeld, GSK, etc. (per official site and Wikipedia) 12

See Also

References

Official Sites

Core Papers

  • Akkaya et al. (2019). “Solving Rubik’s Cube with a Robot Hand.” arXiv:1910.07113. Paper Link 6
  • Andrychowicz et al. (2020). “Learning Dexterous In-Hand Manipulation.” The International Journal of Robotics Research. arXiv:1808.00177 5

Footnotes

Footnotes

  1. Shadow Robot Company Official Site (https://shadowrobot.com/) ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  2. Shadow Hand Wikipedia and company history (https://en.wikipedia.org/wiki/Shadow_Hand) ↩ ↩2 ↩3 ↩4 ↩5

  3. Shadow Dexterous Hand Technical Specification (https://www.shadowrobot.com/wp-content/uploads/2024/06/20240610-UPDATED-shadow_dexterous_hand_e_technical_specification.pdf) ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15

  4. Shadow Robot Blog: How Much Does a Robot Hand Cost? (https://shadowrobot.com/blog/how-much-does-a-robot-hand-cost/) ↩ ↩2 ↩3 ↩4

  5. Andrychowicz et al. (2020). “Learning Dexterous In-Hand Manipulation.” arXiv:1808.00177 ↩ ↩2 ↩3

  6. Akkaya et al. (2019). “Solving Rubik’s Cube with a Robot Hand.” arXiv:1910.07113 ↩ ↩2 ↩3

  7. SynTouch BioTac sensor integration information (https://shadowrobot.com/news/company-news/biotac-sensors/) ↩ ↩2

  8. Shadow DEX-EE Series Official Page (https://shadowrobot.com/dex-ee_series/) ↩ ↩2 ↩3 ↩4 ↩5

  9. “Grasping in the Dark: Compliant Grasping using Shadow Dexterous Hand and BioTac Tactile Sensor” (ICRA 2021) ↩

  10. Google Brain Deep Dynamics Models research (https://shadowrobot.com/blog/machine-learning-innovation/) ↩