Camera (machine vision)

Imagine pointing a torch at a wall in a dark room. You can see the wall perfectly, but you only know it exists in that one spot where the beam lands. Now imagine replacing the torch with a million tiny buckets, each one catching however many photons hit it in a fraction of a second. That grid of light-measuring buckets is the heart of every digital camera — in your phone, on a security system, or mounted to a robot's head.

That grid is called an image sensor, and the science of using it to give machines useful visual information is called machine vision.

How a camera sensor works

The most common image sensor today is the CMOS sensor (Complementary Metal-Oxide Semiconductor). It is a flat chip covered in millions of tiny photodiodes — components that generate a small electrical charge when struck by light. Each photodiode corresponds to one pixel (picture element). The brighter the light hitting a photodiode, the higher the charge it produces. Read all those charges out at once and you have a grid of brightness values: an image.

Colour cameras add a Bayer filter — a pattern of red, green, and blue micro-lenses placed over pairs of pixels. By comparing adjacent R, G, and B readings, the camera's processor reconstructs colour at every point.

A robot's software then takes that grid of numbers and runs algorithms to answer questions: Is there an object here? How far away is it? What is it? This is the field of computer vision, which is distinct from the hardware camera itself but inseparable from it in practice.

A real-world example

Boston Dynamics' Spot robot uses five stereo camera pairs covering 360 degrees. Each pair combines two cameras offset by a fixed distance, much like human eyes. By comparing how the same point appears slightly differently in each camera, the robot's software triangulates distance — a technique called stereo vision. Spot uses this to avoid obstacles and climb stairs without a lidar in sight.

Why it matters

Cameras are by far the cheapest and most information-dense sensors available to robots. A ₹1,000 webcam produces millions of data points per second; a lidar of equivalent spatial resolution costs orders of magnitude more. That density is also the challenge: processing a 1080p image 30 times per second is computationally expensive, and making sense of it reliably in changing lighting, rain, or cluttered environments remains one of robotics' hardest unsolved problems.

Without cameras, a robot cannot read labels, recognise faces, track moving objects, or understand the fine detail of its environment. Every major autonomous system — self-driving cars, surgical robots, warehouse drones — treats camera data as its primary source of truth about the visual world.

If cameras see colour and texture but struggle with distance, what happens when you give a robot two cameras spaced apart like a pair of eyes?