Sense, plan, move, repeat

Think about picking up a glass of water. Your eyes find it. You decide where your fingers will go. Your hand moves, and all the way there you keep checking and correcting.

A robot works in the same kind of loop. It senses the world through cameras and other sensors. It plans what to do next. It moves a little. Then it senses again, because the world may have changed, and the loop goes round again and again.

The loop matters because plans rarely stay right for long once a hand meets real objects. A cup gets nudged. A box is heavier than it looked. The robot only copes if it notices and adjusts.

Seeing isn’t holding

A camera can tell a robot where an object is and roughly what shape it has. It cannot tell the robot how heavy the object is, how slippery its surface is, or whether it will squash. A full carton and an empty one look the same.

Those hidden facts decide how a grip should go. Squeeze too softly and the object slips. Squeeze too hard and it crushes. You solve this without thinking, because your fingertips feel the first tiny slide and tighten before anything falls.

Giving robots a sense of touch

Researchers build touch sensors to give robots some of that feeling. One well-known design is called GelSight. It uses a soft rubbery slab covered with a reflective skin. When the slab presses on something, the way its skin dents and stretches reveals the shape of what it touches. In a 2017 paper, the researchers tested it with a robot arm and 37 everyday objects, and found it could usually detect when an object was sliding or twisting in the grip.

A hand that works without eyes

A robot can also sense its own body. In September 2026, researchers at ETH Zurich described a five-fingered robot hand that walks on its fingers.

Balancing on four fingers, it used the fifth to press arrow keys on a keyboard, with no camera feedback at all. Its controller reads the hand’s joint angles, how it is tilted and how fast it is turning, plus which key a person has asked for. A person lined the keyboard up with the hand before the test. After that, it pressed the correct key 29 times out of 32.

The paper is a preprint and has not yet been peer reviewed. It still makes the point well. Cameras are only one way to sense, and a robot that feels its own body can do some jobs with no vision at all.

In one sentence: a robot senses, plans and moves in a fast loop, and seeing an object tells it far less than touch does about how to hold it.