Height Scan

Samples ground height around a robot in a fixed pattern, used as terrain input for a locomotion policy.

A height scan is a commonly used terrain observation in reinforcement-learning locomotion for legged robots: centered on the body, or around each foot, it samples ground height at a number of points in a fixed pattern, then stacks these into a vector fed to the policy network. In simulation, this is usually done by casting rays straight down from above the robot and reading off ground-truth height directly; a real robot has no such bird’s-eye view, so it first has to build an elevation map (a 2.5D grid of ground height) with a depth camera or lidar, then sample from that map. A real elevation map has noise and occlusion, so a common approach trains a teacher policy on noise-free height first, then distills it into a student policy that reads the real, noisy elevation map. A policy that relies only on proprioception, with no height scan, is called blind locomotion.

ExampleIsaac Lab’s velocity-tracking task mounts a ray-casting sensor on the body, measuring ground height on a 1.6 m × 1.0 m grid at 0.1 m spacing; Miki and colleagues’ 2022 ANYmal work instead samples 52 points in 5 concentric rings around each foot, for 208 dimensions across all four legs.

Also called

height_scan, Terrain Height Map, Height Samples

Related

Elevation Map · Perceptive Locomotion · Blind Locomotion · Teacher-Student Distillation · NVIDIA Isaac Lab · legged_gym

Sources

Isaac Lab velocity_env_cfg.py（height_scanner 与 height_scan 观测定义） (Chinese)
Learning robust perceptive locomotion for quadrupedal robots in the wild（Science Robotics 2022）
legged_gym legged_robot_config.py（measured_points 配置） (Chinese)