Occupancy Networks: Rethinking How Cars See

Bounding boxes are outdated. Occupancy networks predict whether each 3D voxel is free, occupied, or unknown — capturing arbitrary shapes like debris, strollers, or open doors.

From Boxes to Voxels

Tesla's Occupancy Network outputs a 3D grid at 10cm resolution from 8 cameras. Waymo's version fuses LiDAR supervision. Result: the car sees a pile of boxes as occupied space, not "unknown".

This unlocks handling of long-tail objects that never appeared in training labels.