Heightfield Terrain

Representing uneven terrain with a 2D grid that stores the ground height at each point — the standard approach for legged-robot training.

A heightfield is a 2D elevation matrix: the ground is divided into a grid, and each cell stores just one height value, much like a grayscale image where lighter means higher. Simulators use it as a cheap way to represent ramps, stairs, rubble, and other uneven terrain. MuJoCo's hfield asset can be loaded from a PNG or binary file, or generated at runtime from sensor data, and is treated as a set of triangular prisms during collision detection; Isaac Lab instead converts a heightfield into a triangle mesh, and comes with built-in terrains such as random rough ground, pyramid slopes, stairs, discrete obstacles, waves, and stepping stones, each with a difficulty parameter, making it easy to train from easy to hard using a terrain curriculum. The limitation is that a heightfield can only have one height per horizontal position, so it can't represent overhangs or caves.

ExampleIsaac Lab's pyramid stairs terrain has taller steps as its difficulty parameter increases; when training quadruped locomotion, it's often mixed with randomly rough terrain and difficulty is ramped up gradually via a terrain curriculum.

Also called

Height Map Terrain, hfield

Related

Terrain Curriculum · Rough-terrain Locomotion · Elevation Map · Height Scan · legged_gym · MuJoCo (Multi-Joint dynamics with Contact)

Sources

MuJoCo XML Reference: asset/hfield
Isaac Lab API: isaaclab.terrains