Tactile Sensors

Overview

Tactile sensors emulate the touch functionality of human skin, providing information about contact area force distribution, geometry, texture, temperature, and more. Unlike 6-axis F/T sensors that measure "total force," tactile sensors measure spatially distributed contact information.

The human fingertip has approximately 240 tactile receptors per cm² with a resolution of about 1 mm. Modern tactile sensors are gradually approaching this level.

Vision-Based Tactile Sensors

GelSight (MIT)

GelSight is the most influential vision-based tactile sensor, proposed by Edward Adelson's team at MIT in 2009.

Working Principle:

A transparent elastomer surface is coated with a reflective layer
An object contacts the elastomer, deforming its surface
Internal LEDs illuminate the deformed surface from multiple directions
A miniature camera captures the deformation image
Photometric stereo reconstructs the 3D surface geometry

Photometric Stereo Principle:

From $k$ illumination images $I_1, I_2, \ldots, I_k$ taken from different lighting directions, recover surface normals:

\[I_k(x, y) = \rho(x, y) \cdot (\vec{n}(x, y) \cdot \vec{l}_k)\]

where $\rho$ is the albedo, $\vec{n}$ is the surface normal, and $\vec{l}_k$ is the $k$-th light source direction.

Three light sources suffice to solve for the normal $\vec{n} = (n_x, n_y, n_z)$, and integration yields the depth map.

Typical Specifications:

Parameter	Specification
Spatial Resolution	~25 μm
Force Resolution	~0.01 N
Contact Area	~14 × 10 mm
Frame Rate	30-60 fps
Depth Resolution	~2 μm

Output Data:

RGB tactile image (640×480 typical)
3D height map (reconstructed via algorithms)
Contact region mask
Normal/tangential force estimates

DIGIT (Meta AI)

DIGIT is a compact vision-based tactile sensor developed by Meta (formerly Facebook) AI Research, inspired by GelSight.

Improvements:

Smaller: ~20 × 27 mm contact area, suitable for mounting on robot fingers
Open source: Hardware design and software are fully open source
Low cost: Material cost ~$15 (even lower for mass production)
Standardized: Unified interface for community reproducibility

DIGIT Structure
┌─────────────────┐
│   Elastomer Gel  │ ← Contact surface
├─────────────────┤
│  Reflective Coat │
├─────────────────┤
│  LED Lighting    │
│     (RGB)        │
├─────────────────┤
│  Miniature Camera│ ← OV5640 (USB)
├─────────────────┤
│  PCB + Connector │
└─────────────────┘

DIGIT Ecosystem:

PyTouch: Meta's open-source tactile perception library
TACTO: Tactile sensor simulator (based on PyBullet)
Perception tasks: Contact detection, force estimation, slip detection, object classification

GelSlim

A thin-profile vision-based tactile sensor developed by the MIT team:

Only 7 mm thick (GelSight ~25 mm)
Uses side illumination + light guides
Suitable for integration into parallel grippers

9DTact

Proposed by Southern University of Science and Technology and others:

Uses a fisheye lens instead of a planar camera
Can measure 3D contact force distribution
Spherical contact surface, suitable for fingertip form factor

BioTac (SynTouch)

BioTac emulates the multi-modal sensing capabilities of the human fingertip:

Sensing Modalities:

Modality	Sensor	Information
Force	19 impedance electrodes	Spatial force distribution
Vibration	1 pressure sensor (AC)	Slip/Texture (up to 1 kHz)
Temperature	1 thermistor	Object material identification
Static Pressure	1 pressure sensor (DC)	Total contact force

Structure:

Rigid skeleton + elastic skin
Conductive liquid fills the space between skin and skeleton
Fingerprint-like surface texture enhances friction and vibration sensing

Applications:

Object material classification (metal/wood/plastic/fabric)
Fine manipulation (page turning, USB insertion, etc.)
Prosthetic tactile feedback research

Limitations:

Expensive (~$5000 each)
Discontinued (after SynTouch was acquired)
Proprietary data interface

ReSkin (Meta AI)

ReSkin is a magnetic thin-film tactile skin:

Working Principle:

Magnetic particles embedded in an elastomer
Magnetometer array on the bottom PCB
Contact deforms the elastomer, moving magnetic particles, changing the magnetic field
Magnetometers detect field changes and infer force

\[\vec{B}_{measured} = f(\vec{F}_{contact}, \text{geometry})\]

Specifications:

Parameter	Specification
Thickness	~3 mm
Sampling Rate	400 Hz
Force Range	0.1 ~ 10 N
3-axis Force	Supported (normal + tangential)
Replaceable	Elastomer is detachable and replaceable
Cost	~$5/piece

Advantage: Cheap, thin, and replaceable magnetic skin (consumable approach)

Tactile Skin Arrays

Large-Area Tactile Coverage

Humanoid and service robots require whole-body tactile perception.

Capacitive Tactile Skin:

\[C_{ij} = \varepsilon_0 \varepsilon_r \frac{A}{d_{ij}}\]

Each tactile unit (taxel) is a miniature capacitor; when force is applied, plate spacing $d$ decreases and capacitance increases.

Typical Solutions:

Solution	Taxels per Unit Area	Sampling Rate	Features
Shadow Robot iCub skin	~60/dm²	50 Hz	Triangular modules
Bosch skin	~16/dm²	100 Hz	Flexible PCB
Roboskin (EU)	~12/dm²	25 Hz	Modular triangles

Piezoresistive Tactile Arrays:

Use conductive rubber or conductive fabric
Resistance decreases under force
Low cost but lower accuracy and consistency

Tactile Gloves

Used for human hand motion capture (teleoperation data collection):

Manus VR Glove: Finger bending + haptic feedback
StretchSense: Capacitive stretch sensing
DIY approach: Piezoresistive fabric sensors + Arduino

Resolution and Sensitivity Comparison

Sensor	Spatial Resolution	Force Sensitivity	Frame Rate	Modalities	Price
GelSight	~25 μm	~0.01 N	30 fps	Geometry+Force	~$200 DIY
DIGIT	~50 μm	~0.03 N	60 fps	Geometry+Force	~$15 DIY
BioTac	~1 mm	~0.01 N	100 Hz	Force+Vibration+Temp	~$5000
ReSkin	~5 mm	~0.1 N	400 Hz	3-axis force	~$5
Capacitive Array	~5-10 mm	~0.1 N	50-100 Hz	Normal force	Medium

Selection Recommendations:

Fine manipulation research -> GelSight / DIGIT (high-resolution geometry)
Whole-body tactile -> ReSkin / Capacitive array (large-area coverage)
Multi-modal perception -> BioTac (discontinued, consider alternatives)
Rapid prototyping -> DIGIT (open source, low cost)

Tactile Data Processing

Contact Detection

The most fundamental task: determine whether contact exists

\[\text{contact} = \begin{cases} 1 & \text{if } \|I_{current} - I_{baseline}\| > \theta \\ 0 & \text{otherwise} \end{cases}\]

Force Estimation

Estimate contact force from tactile images/signals (typically requires neural networks):

\[\hat{F} = f_{NN}(I_{tactile}; \theta)\]

Training data source: Simultaneously captured tactile images and 6-axis F/T sensor ground truth.

Slip Detection

Detect slip through temporal changes in tactile signals:

Optical flow: Inter-frame pixel displacement in consecutive tactile images
Vibration spectrum: Slip generates high-frequency vibration signals
Contact area change: Contact region shape changes during slip

Object Recognition

Tactile sensing can distinguish properties that are difficult for vision:

Hardness (rigid/soft)
Texture (smooth/rough)
Material (metal feels cold/wood feels warm)
Weight (through grasping force feedback)

Simulation Environments

Simulator	Supported Sensors	Physics Engine	Open Source
TACTO	DIGIT, OmniTact	PyBullet	Yes
Taxim	GelSight	FEM	Yes
Isaac Gym	General tactile	PhysX	Yes

Simulation is an important complement to tactile learning -- real tactile data collection is slow and expensive.

Dexterous Hands -- An important platform for tactile sensors
Force Sensing Overview -- A holistic view of force sensing
6-Axis Force/Torque Sensor -- The gold standard of F/T sensing

References

Yuan, W. et al., "GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force," Sensors, 2017
Lambeta, M. et al., "DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation," RA-L, 2020
Bhirangi, R. et al., "ReSkin: versatile, replaceable, lasting tactile skins," CoRL, 2021
Wettels, N. et al., "Biomimetic Tactile Sensor Array," Advanced Robotics, 2008