THE MACHINE THRESHOLD
Robotics / EXPLAINER / 2 MIN READ + OPTIONAL DEEP DIVE

A robot finger can see where it feels

A curved touch sensor uses mirrors to let one internal camera watch a long contact surface. The optical design expands where a robot can sense a grasp.

AI-assisted synthesis · Published 2026-09-28 · Updated & sources checked 2026-09-28
How we research and correct our work

A camera does not have to face outward to help a robot feel.

Looking from inside the skin

In MIT’s October 2023 account of GelSight Svelte, an internal camera observes a soft skin as contact changes its shape. Two mirrors extend the camera’s view along a curved finger. The useful idea is that touch becomes something a camera can measure, over more than a small fingertip. [1]

Concept cutaway of a curved blue tactile finger touching a ridged object, with an internal optical path suggested.
AI-generated concept, not a research photograph or an exact GelSight Svelte schematic.

Folding a view into a narrow space

Zhao and Adelson’s September 2023 paper describes one flat mirror and one curved mirror, with a camera housed inside the finger. The folded optical path addresses a packaging problem: extending the sensing surface without putting a bulky camera outside it. A flexible backbone also bends and twists; tracking its deformation supports estimates of the torques acting on the finger. [2]

Sensing is one part of grasping

The paper reports object-holding demonstrations with three fingers, including a small brick, screwdriver and flashlight. Those grasp tasks were open loop: the demonstration did not use tactile feedback to continually adjust the grasp. The result therefore supports a sensor design and its measurements, rather than a claim that a robot can autonomously handle arbitrary objects. [2]

Go a little deeper

Optional reading · about 1 more minute

Two questions inside one grasp

Hypothetical example: Imagine holding a ridged tool. A contact image could help describe where its ridges press, while a bending estimate could describe loading on the finger. Those are different questions. Deciding how the hand should move next would require a controller that uses the information.

What to look for in a follow-up

Our interpretation: A persuasive next demonstration would show the hand responding to changing contact, with clear failure cases and repeated trials. Broader sensing coverage creates an opportunity for control; it does not by itself establish reliable manipulation.

Original sources

Attributed synthesis, not original reporting. Examples labeled hypothetical or illustrative are explanatory. Reviewing a source does not independently validate its findings.

  1. MIT: finger-shaped sensor enables more dexterous robots ↗

    October 4, 2023 institutional report; Mirror mirror mechanism read September 28. Historical research, not a current product announcement.

  2. Zhao and Adelson: GelSight Svelte ↗

    September 19, 2023 v1 paper; sections II–IV, optical design and open-loop object holding read September 28. No quantitative performance or universal grasp claim adopted.

Suggest a correction

Know someone who would find this interesting?

Share this story on Facebook ↗ ·

Follow on Facebook ↗ for story highlights and questions to explore next.

Where this question leads next

Follow new explainers and updates →