Research · Perception · Machine touch
Vision and hearing are solved; touch is the sense that gates dexterity. At the moment of a grasp the hand hides the object, and force, slip and compliance are not optical quantities — so a robot that can see still can't reliably hold. A tactile skin supplies what an outside camera cannot. Touch is not one problem, though: covering an arm and reading a fingertip are different jobs, so we focus on the three approaches that are actually viable.
The choice is set by the job, not by a list of physics. A large area — an arm or palm — needs thin, cheap, flexible skin. A fingertip doing fine manipulation needs the richest possible contact data. These are the three approaches with a real path to deployment; it is not an exhaustive list.
Maturity — deployable today · Opportunity — untapped headroom. Ordinal 1–5, our synthesis of the cited reviews — not a measured index.
Why a camera counts as touch. A vision-based skin never looks at the object — it films the deformation of a soft gel pressed against it. The quantity measured is mechanical force, not light from the scene; the camera is a high-resolution readout, not an eye. That is the whole point: at the moment of grasp the hand hides the object from any outside camera, so the sensor has to live at the contact itself. It confirms vision's blind spot rather than curing it.
Every tactile system is the same loop. The six approaches are interchangeable skins feeding one perceive-then-act backbone — which is why the contact data and control policy, not the sensor, carry the value.
A closed reflex loop. Unlike a camera watching from outside, touch closes a control loop in milliseconds — the hand feels a slip and tightens its grip before vision could even register it. And at the instant of contact the hand occludes the object, so touch is the only sense still reporting.