Occupancy grid

An occupancy grid carves the world into a regular lattice of cells, two-dimensional for floor-plan navigation or three-dimensional for full volumes, and assigns each cell a probability of being occupied. Cells the sensors have seen through become confidently free, cells where returns land become confidently occupied, and unobserved space stays uncertain. The result is a simple, probabilistic picture of where a robot can and cannot go.

It is one of the foundational representations in classical robotics, and it remains widely used because planners consume it directly.

As range measurements arrive from LiDAR, depth cameras, or sonar, each ray updates the cells it passes through, lowering their occupancy probability, and the cell where it terminates, raising it. Bayesian updates accumulate evidence across many readings, so noise averages out over time. Resolution is the key trade-off: finer cells capture more detail but cost more memory and computation.

Occupancy grids are the classical answer to the question every mobile robot asks, where is it safe to move, and they remain the bridge between raw sensing and planning in countless deployed systems. They are also the conceptual ancestor of learned occupancy networks, which predict occupancy directly with neural networks rather than accumulating it ray by ray.

An occupancy grid accumulates occupancy probabilities from sensor rays using explicit Bayesian updates, while an occupancy network learns to predict occupancy directly from sensor input with a neural network.

Cells no sensor ray has observed, which stay at their prior probability. Distinguishing unknown from free is important for safe planning.

Occupancy network, SLAM (simultaneous localization and mapping), Motion planning, LiDAR, Point cloud.

Learn: What is physical AI? · Physical AI data platform guide

Try it: Build a 3D dataset with Point-E and FiftyOne · FiftyOne 3D visualization