White paper
Tactile sensing technologies, compared
How the main touch-sensing methods work, and the trade-offs that decide which one belongs on your robot.
A grasp mixes several signals at once: how hard you press, how much the object drags sideways as it starts to slip, the tiny vibrations as a surface passes under the skin, the shape of the contact patch, sometimes temperature. Most sensors read two or three of those well and stay blind to the rest.
So choosing a sensor is really choosing which signals you care about, and what you will give up in resolution, speed, durability, or cost to get them. A gel fingertip and a capacitive skin are not two versions of the same thing. They run on different physics, and their raw output has almost nothing in common. Here are the five approaches you are most likely to weigh, and where each one earns its place.
The technologies
Vision-based (optical / gel)
A camera watches a soft skin deform.
A camera sits behind a clear elastomer coated with a reflective skin and lit by colored LEDs. Press something into the gel and the shading reveals the exact shape of the contact, down to fine texture. Printed dots on the skin let you track how the surface drags and slips.
- Resolves surface detail finer than a millimeter, plus texture and edges
- Marker dots make shear and slip visible directly
- It is a camera, so the whole computer-vision toolkit carries over
- The optics need room behind the gel, so fingertips get bulky
- Frame rate caps how fast you can react to a slip
- The soft skin scratches and has to be recast periodically
- Force is inferred, not measured, so calibration is extra work
- Best for
- Dexterous, contact-rich manipulation, and anything where you want to see the contact rather than just feel it.
- Examples
- GelSight, DIGIT, GelSlim, OmniTact
Capacitive
Pressure changes the capacitance between electrodes.
Two layers of electrodes sandwich a squishy dielectric. Press down and the gap shrinks, which changes the capacitance at that spot. A finger hovering nearby shifts it too, before anything touches. It is the same effect behind phone touchscreens, laid out as a grid.
- Thin and bendy, so it wraps around curved parts as a skin
- Picks up the lightest touch, and senses a hand before contact
- Cheap, low power, and a mature technology
- Drifts with humidity and picks up noise from nearby motors
- Neighboring cells bleed into each other in dense arrays
- Getting an absolute force number out of it is hard
- Best for
- Large-area skins on arms and shells, proximity, and safe human contact.
- Examples
- Touchscreen-derived arrays, many research e-skins
Piezoresistive (resistive / FSR)
Squeezing the material lowers its resistance.
A conductive rubber or polymer drops in resistance as you compress it. Wire a grid of them and each cell reports how hard it is being pressed through a simple voltage reading. The cheap hobby version is the force-sensitive resistor.
- About as simple and cheap as touch sensing gets
- Easy to build into dense pressure-map arrays
- Plenty of off-the-shelf parts
- Creep: hold a steady load and the reading keeps drifting
- The response is nonlinear and the usable range is narrow
- Readings wander with temperature
- Best for
- Cheap pressure maps, and grippers that mostly need to know whether, and how hard, they are touching.
- Examples
- FSRs, conductive-elastomer taxel skins
Magnetic (Hall-effect)
A magnet in soft skin moves over Hall sensors.
Embed a small magnet, or a film of magnetized particles, in a rubber skin above a few Hall sensors. Load pushes the magnet sideways or down, the field shifts, and you can recover how the skin moved in all three directions. That gets you shear, not just push, with no wires running through the contact surface.
- Reads three-axis force, so it catches the sideways drag that signals a slip
- Nothing fragile at the contact surface, so it takes a beating
- Fast enough to catch quick events
- A nearby magnet or motor throws it off
- The axes couple together, which makes calibration fiddly
- You get fewer sensing points than a camera or a dense array
- Best for
- Slip detection, and fingertips that have to survive real, repeated grasping.
- Examples
- ReSkin, magnetic-tip fingertips
Piezoelectric
Straining the material makes a charge.
A piezoelectric film such as PVDF puts out a charge when it is strained, and the faster it strains the bigger the signal. Run that through a charge amplifier and it behaves much like an accelerometer would: quick to catch the buzz of a passing texture, or the first micro-slip of an object starting to go.
- Very fast, so it catches vibration and the instant a slip begins
- Generates its own signal, with nothing to power at the sensor
- Thin and light
- Only feels change: hold a constant force and the signal fades to nothing
- The charge leaks away, so steady and slow loads are lost
- Needs a fussy charge amplifier up front
- Best for
- Slip and vibration sensing, texture ID, and event-driven touch, usually paired with a sensor that handles steady force.
- Examples
- PVDF skins, piezoelectric slip sensors
How to choose
Where they trade off.
- Force axes
- Some read only straight-down pressure (piezoresistive). Others also read the sideways drag that predicts slip (magnetic, marker-based vision).
- Static vs dynamic
- Piezoelectric only feels change. The others hold a steady load, though some creep while they do.
- Spatial resolution
- Vision is in a class of its own. Magnetic trades detail for toughness.
- Bandwidth
- Piezoelectric and magnetic react fast. Vision is stuck at camera speed.
- Robustness
- Sealed magnetic parts shrug off abuse. Gels wear, and capacitive skins drift.
- Cost
- Piezoresistive is the cheapest to build. Vision needs optics and a chunk of compute.
In closing
Why a common representation matters
Because the raw signals have so little in common, code rarely moves between sensors. A model trained on gel images has learned pixels. One trained on a magnetic fingertip has learned a handful of field readings. Swap the hardware and you are mostly starting over, which is why every lab keeps rebuilding the same drivers, calibration, and file formats.
That is the problem Contour works on: one way to write down a moment of contact, the reading, its geometry, and where the sensor sat, so the data looks the same whatever produced it. Pick a sensor for hardware reasons, and your datasets and models come along with it.