What is an occupancy network?

An occupancy network predicts the 3D structure of a scene by asking, for many points in space, whether each one is filled with something solid. Instead of drawing boxes around known object categories, it produces a dense occupancy grid or field that captures arbitrary shapes, including obstacles that do not fit any predefined class.

This class-agnostic view is valuable for driving and robotics, where anything solid matters whether or not the system has a label for it.

Key takeaways

It predicts whether each point in 3D space is occupied.
The output is dense and does not rely on predefined classes.
It captures arbitrary obstacles that box-based detection can miss.

How it works

The network takes sensor input, often multiple camera views or LiDAR, and predicts occupancy over a 3D volume around the vehicle or robot, sometimes with a semantic label and motion estimate per voxel. Many systems build this representation in a bird's-eye-view space that fuses several sensors into a common frame. The occupancy volume then feeds planning and collision avoidance.

Why it matters

Object detectors can miss unusual obstacles they have no category for, and occupancy networks close that gap by representing space itself rather than named objects. That makes them increasingly central to autonomous driving perception, complementing detection with a dense, general model of what is solid.

Frequently asked questions

How is an occupancy network different from object detection?

Object detection labels known categories with boxes, while an occupancy network predicts whether space is filled, capturing arbitrary shapes without needing a class.

Why is occupancy useful for driving?

It flags any solid obstacle, including unusual ones a detector was never trained on, which is critical for safe collision avoidance.

Jesse Mostipak is the SEO and Content Manager at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn

Occupancy network

Building visual or physical AI? 
Let's talk.