point cloud

A point cloud is simply a big bag of dots scattered in 3D space, where each dot marks one spot the robot has measured on the surface of something real. Imagine spray-painting a single glowing speck onto every place your finger touches an object — the chair, the floor, the coffee mug — and then floating all those specks in the air with nothing connecting them. Step back and the cloud of specks traces out the shape of the room and everything in it, even though it is just a list of points, not a solid model. Each point carries at least three numbers: how far left or right, how far up or down, and how near or far it sits, measured from the sensor.

Robots get point clouds from depth sensors. A lidar sweeps a laser around and times how long each pulse takes to bounce back, turning those times into distances and then into dots; a depth camera does much the same using infrared light. The result is the raw, unprocessed picture of geometry the robot starts from — there are no labelled objects yet, no surfaces, just thousands or millions of bare dots. Points can also carry extra data per dot, such as color borrowed from a camera or a brightness value, so a cloud can look like a ghostly 3D photograph of a scene.

Because a point cloud is raw and unstructured — the dots have no order, no built-in notion of which belongs to the table and which to the cup — most of the work in perception is turning it into something meaningful: smoothing out sensor jitter, stitching several scans together, grouping dots into surfaces, or matching the cloud against a known shape. It is the common starting material for mapping a building, recognizing an object, or letting a self-driving car understand the road around it.