Elevation Map

• Also called: height map, 2.5D map, digital elevation model
• Category: Sensing › 3D maps & reconstruction
• Tags: mapping, terrain, legged-locomotion, 3d-perception, navigation

A 2.5D grid map of terrain that stores one height value, often with an uncertainty, for each horizontal cell; a standard input for legged and off-road robot planning.

Definition

An elevation map is a grid over the ground plane in which each cell holds the height of the terrain surface at that location. Because it stores only one height per cell, it is often called a 2.5D representation: richer than a flat occupancy grid, and much cheaper than a full 3D voxel map. Cells usually also carry a variance or confidence, and many systems add further layers for surface normals, slope, roughness or traversability computed from the heights.

Building an elevation map means fusing range data over time. Each depth image or lidar scan is transformed into the map frame using the robot's pose, the points are binned into cells, and each new height measurement is merged with the stored estimate, often with a Kalman-filter-style update weighted by sensor noise. Errors in the pose estimate smear the map, so odometry drift directly limits map quality. Robot-centric elevation mapping, developed at ETH Zurich for legged robots such as ANYmal, keeps the map attached to the robot and increases the uncertainty of cells as the robot moves away from where they were observed, so that terrain near the feet stays consistent even when the global position drifts. GPU implementations make it possible to update the map at camera rate.

The single-height assumption is also the main limitation. Overhangs, tables, tunnels and multi-floor structures cannot be represented, and areas hidden behind steps or rocks remain unknown unless they are filled by interpolation or learned inpainting. Multi-level surface maps and full 3D maps handle these cases at extra cost. For legged locomotion, elevation maps feed foothold planners, provide height samples around each foot as inputs to learned policies, and support traversability estimates for path planning.

Why it matters

Legged robots and off-road vehicles need to know where the ground is, not only where obstacles are. An elevation map provides that in a compact form that planners and learned policies can query quickly, and it has become a common interface between perception and terrain-aware locomotion.

Key applications

• Perceptive locomotion for quadrupeds and humanoids on stairs and rough ground
• Foothold planning and step-edge avoidance
• Traversability analysis for off-road vehicles and planetary rovers
• Height-scan inputs for learned locomotion policies
• Terrain mapping for agricultural and construction robots

In practice

On a quadruped climbing stairs, depth cameras on the body produce point clouds several times per second. Each cloud is transformed into a robot-centered map frame with the pose from visual-inertial or leg odometry and fused into a grid with cells a few centimeters across. The controller samples heights in a pattern around each foot and feeds them, together with proprioception, to a locomotion policy, or passes the map to a foothold planner that rejects cells near step edges. Engineers watch for artifacts such as spikes caused by the robot's own legs entering the camera view and filter them out before the map reaches the controller.

Common misconceptions

• An elevation map is a full 3D map — it stores one height per cell and cannot represent overhangs, tables or multiple floors
• Cells without data are flat ground — unobserved cells, often behind obstacles, are unknown and should be treated with caution or filled explicitly
• Map quality depends only on the depth sensor — pose errors smear heights across cells, so odometry quality matters as much as the sensor

Who should know this

• Legged locomotion engineers working on perceptive control
• Field and off-road robotics developers
• Mapping and perception engineers
• Planetary and agricultural robotics researchers

Related technologies

• Grid Map Library
• Robot-Centric Elevation Mapping
• Visual-Inertial Odometry
• OctoMap
• Multi-Level Surface Maps

Related terms

• Terrain Traversability
• Point Cloud
• Depth Camera
• Odometry
• Octree
• Foothold Planning

Related projects

• LingBot-Map: Streaming 3D reconstruction from one camera: pose and depth for every frame, live
• Cartographer: Google's real-time 2D and 3D SLAM system for multi-platform robotics
• OpenDroneMap: Open-source photogrammetry toolkit that turns drone images into maps and 3D models
• Nav2: ROS 2 navigation stack with path planning and obstacle avoidance for mobile robots
• DimOS: Python robotics runtime with agents, SLAM and teleop, no ROS required
• RTAB-Map: Versatile graph-based SLAM supporting RGB-D, stereo, and LiDAR with real-time operation