Voxel

A small cube in 3D space — the volumetric equivalent of a pixel.

A voxel (a blend of “volume” and “pixel”) is one small cube produced by slicing 3D space into a regular grid; each cell can store whether it’s occupied, a color, a distance to the nearest surface, or a feature vector, just as a pixel does in a 2D image. Compared with a point cloud — a scattered set of 3D points — voxels are arranged regularly, so they can be processed directly with 3D convolutions and make it easy to check whether a given location is blocked; the tradeoff is that memory use grows fast with resolution, since halving the cell size multiplies the cell count by eight. Common robotics uses include occupancy grid maps and octree maps, which use voxels to mark where a robot can go; the truncated signed distance function (TSDF), which stores a distance value in each voxel for 3D reconstruction; and voxel downsampling of point clouds, which merges all points inside one cell into a single average point. The manipulation policy PerAct divides the workspace into a 100×100×100 voxel grid and directly predicts which cell the gripper should move to next.

ExampleIn Open3D, a single line — pcd.voxel_down_sample(voxel_size=0.05) — downsamples a dense point cloud to one point per 5-centimeter cell (with coordinates in meters), sharply cutting the point count before it’s sent on for registration or grasp detection.

Also called

Voxel Grid, Voxel Downsampling, Volumetric Pixel

Related

Point Cloud · Occupancy Grid Map · OctoMap · Truncated Signed Distance Function · Point Cloud Filtering & Voxel Downsampling · PerAct

Sources

Wikipedia: Voxel
Open3D Tutorial: Point cloud (Voxel downsampling)
PerAct: Perceiver-Actor project page