The tactile sensors giving robot hands a sense of touch

A camera can watch a thousand robots and never wear out. A tactile sensor gets one hand, feels every object that hand ever grasps, and slowly dies from the contact. Peel the fingertip off a research robot after a few weeks of picking and placing, and you often find a scuffed gel, a torn membrane, or an electrode that has quietly drifted out of calibration. That asymmetry, cheap immortal eyes against fragile mortal skin, is why robot touch is still a hardware problem before it is a data one.

Vision converged on one device. Almost every camera on Earth turns light into a grid of numbers the same way, and any two of them speak a format called an image. Touch never had that moment. Ask five labs how their robot feels a screw thread and you get five physically different answers: a sheet that changes resistance, a chip that measures a squeezed pocket of air, a gel photographed from the inside. None of them dominates, and the differences run all the way down to the physics.

Contact is not one quantity. Press a fingertip on a table and there is normal force pushing straight in, shear dragging sideways, the tiny slip that precedes a drop, high-frequency vibration that reads as texture, and sometimes heat flowing across the boundary. Every tactile technology is a bet on which of those to capture cleanly and which to give up. Here is what the main families measure, what they cost you, and why none of it has produced a large shared dataset the way cameras did.

Turning a squeeze into a number

Most robot skins are electrical. They tile a surface with taxels, tactile pixels, and read some property that changes when the surface is pressed.

Resistive and piezoresistive skins are the workhorses. A pressed material conducts differently, and a grid of such junctions gives a coarse pressure map for very little money. They go thin, they wrap around curved fingers, and they scale to large areas cheaply. The price is signal quality: they drift, they show hysteresis (the reading on the way down does not match the way up), and they are poor at telling shear from normal load.

Capacitive arrays measure the gap between two conductive layers through a soft dielectric. Squeeze the dielectric and the capacitance rises. This is the technology under a phone screen, shrunk and stiffened for fingertips, and it is sensitive and low-power. Pack the taxels close for resolution and neighboring cells start to cross-talk, and the wiring count climbs fast, one reason dense capacitive skins are hard to run across a whole hand.

Barometric and MEMS designs hide a mass-produced pressure die under a dome of rubber. Force squeezes the trapped fluid or air, the barometer chip reads it, and you inherit the ruggedness and low cost of a part made by the billion for phones and cars. One die is one taxel, though, so spatial resolution is coarse unless you flood the hand with chips and the wiring that comes with them.

Piezoelectric and magnetic approaches fill the gaps. Piezoelectric films generate charge from changing pressure, which makes them excellent at the fast stuff, vibration, slip onset, texture, and useless at holding a steady weight, because a constant force produces no changing charge. Magnetic skins such as ReSkin suspend magnetic particles in a soft elastomer over Hall-effect sensors, then read shear and normal force from how the field shifts; the elastomer is a cheap consumable you peel off and replace when it wears.

A camera pointed at a gel

The other big family inverts the whole problem. Instead of wiring a surface, put a soft clear gel over a small camera, light it from inside, and film the gel deforming as it presses on the world. GelSight pioneered this, DIGIT made a compact low-cost open version, TacTip tracks a mesh of internal pins, and Toyota Research Institute's soft bubble grippers stretched the same idea over an inflated membrane. The output is not a few numbers. It is a dense tactile image.

That image is the selling point. A photographed gel can resolve surface texture finer than a human fingertip, reconstruct the three-dimensional shape of whatever is pressing on it, and recover shear by watching printed markers slide. It also standardizes on the one sensor the field already knows how to mass-produce and read: a camera. The costs are physical. The gel is bulky, needs a few millimeters of clearance and an optical path, tears and clouds with use, and every frame is a full image, so bandwidth and compute are heavy next to a thin resistive sheet. You trade a wiring problem for an optics-and-wear problem.

What each family can and cannot feel

No single technology wins. Each family is strong on part of the contact signal and blind on the rest, which is why serious hands often carry more than one.

Read down the tradeoff column and a pattern appears. The cheap, tileable, durable options are coarse. The rich, high-resolution options are bulky, fragile, or data-heavy. There is no equivalent of the good-enough webcam that is cheap, standard, and detailed all at once.

Why the corpus never shows up

This is where hardware becomes a data story. Vision scaled because a camera is external, interchangeable, and non-consuming: you bolt the same webcam onto a thousand robots, and every clip lands in one shared format. Touch breaks all three properties at once.

A camera outlives the data it records. A tactile sensor is consumed by it. That single fact, more than any modeling gap, is why there is no ImageNet for touch.

The sensor has to live inside the hand, so you cannot pool data across robots without also standardizing their fingers. Calibration is per-unit and drifts, so a raw reading in sensor counts means nothing to the next lab unless it maps to Newtons. Worst of all, the instrument is destroyed by the very act of measuring: the gel that clouds, the elastomer that abrades, the electrode that fatigues, all change their response part way through a collection run. A vision dataset can be recaptured next year from the same footage. A tactile record is tied to a specific worn sensor on a specific day, and there is still no shared schema for what a touch reading even is across two labs.

So the field has an enormous visual corpus, a growing vision-plus-proprioception corpus in efforts like Open X-Embodiment, and almost no large tactile-manipulation dataset at all. The hardware fragmentation is not a footnote to that scarcity. It is the cause.

Living with fragmentation

The most promising response is to make touch look like vision. Optical sensors such as DIGIT push open, reproducible hardware so that two labs at least share a format, a tactile image, even if their gels wear differently. Open toolkits like Hugging Face's LeRobot give tactile streams somewhere standard to live next to video and actions. Groups working on dexterity, from Physical Intelligence to Toyota Research Institute, treat contact as a first-class channel rather than a bolt-on, and engineer-grade coverage in outlets like IEEE Spectrum tracks a steady stream of new skins out of the robotics literature. None of it has yet produced touch's ImageNet moment.

Give a robot a camera and it inherits a century of imaging and a planet of pixels. Give it a sense of touch and you are still choosing which physics to trust, wiring it into a finger by hand, and watching it wear out as it works. The algorithms are ready to feel. The instruments, and the shared record they would write, are the part that still has to be built one hand at a time.