ouguangjun/Leg-KILO

What it solves

Leg-KILO (also known as kilo-map) provides a real-time LiDAR-based SLAM (Simultaneous Localization and Mapping) system designed to be robust against aggressive platform dynamics, such as those experienced by legged robots. It addresses the challenge of maintaining accurate odometry and global consistency in both structured and unstructured environments while handling high-dynamic motion distortion.

How it works

The system employs a multi-stage approach to fuse sensor data and optimize the map:

• Two-Stage ESKF Frontend: It uses an Error-State Kalman Filter (ESKF) to fuse LiDAR and IMU data. The first stage performs per-point updates to compensate for motion distortion. The second stage uses an Iterated ESKF (IESKF) over the full frame to refine global consistency.
• Hybrid Feature Gaussian Voxel Map: The system maintains a map of uncertain hybrid features (planar and NDT-variant), which improves how features are utilized across different environment types.
• Factor Graph Backend: A Ceres-based factor graph optimization tightly couples loop-closure constraints with odometry factors to reduce global drift. It leverages small_gicp and KISS-Matcher for loop-closure verification.
• Visualization and Storage: It uses the Iridescence library for real-time 3D visualization and allows users to save global maps as single files or tiled blocks.

Who it’s for

This project is intended for robotics developers and researchers working with LiDAR-equipped platforms, particularly those with high-dynamic movements like legged robots, who need robust real-time mapping and localization.

Highlights

• High-Dynamic Robustness: Specifically designed to handle the aggressive motion of legged robots.
• Cross-Platform Support: Compatible with both ROS 1 and ROS 2 (including Melodic, Noetic, Foxy, Humble, and Jazzy).
• Lidar Agnostic: Supports multiple LiDAR models including Velodyne, Ouster, Hesai, and Livox.
• Tightly Coupled Fusion: Combines IMU and LiDAR data via a two-stage ESKF for fast convergence and distortion compensation.