leggedrobotics/elevation_mapping_cupy

What it solves

This project provides a high-performance elevation mapping system for robots, specifically designed for legged robots navigating complex terrains. It solves the problem of creating accurate, real-time 2.5D maps of the environment from point cloud data, while handling common robotic mapping artifacts like height drift and virtual artifacts from overhanging obstacles.

How it works

The system uses GPU acceleration via CuPy to rapidly process large point clouds and perform ray casting for visibility cleanup. It integrates with the Robot Operating System (ROS) and implements a Multi-Modal Elevation Map (MEM) framework. This framework allows the system to fuse diverse data types—such as geometry, RGB information, and semantics—into different layers of the map, which can then be processed by custom plugins to generate additional information like terrain traversability.

Who it’s for

Robotics engineers and researchers focusing on legged locomotion, autonomous navigation, and terrain perception for robots operating in unstructured environments.

Highlights

• GPU-Accelerated Processing: Uses CuPy for real-time efficiency in point cloud registration and ray casting.
• Multi-Modal Integration: Supports fusing geometry, semantics, and RGB data into a layered map structure.
• Drift Compensation: Includes features to tackle state estimation drifts to prevent mapping artifacts.
• Traversability Analysis: Features a learning-based filter to assess if terrain is walkable based on local geometry.
• Artifact Removal: Uses raycasting and exclusion zones to correctly interpret overhanging obstacles and remove virtual artifacts.