sjtuyinjie/Ultra-Fusion

What it solves

Ultra-Fusion addresses the challenge of maintaining reliable localization and mapping (SLAM) in intelligent transportation systems when sensors fail or perform poorly. It specifically targets "corner-case" scenarios such as poor lighting, LiDAR degeneracy (e.g., in long corridors), wheel slippage, and GNSS outages, as well as issues caused by timing delays or incorrect sensor calibration.

How it works

The framework uses a tightly-coupled sliding-window optimizer that processes asynchronous sensor streams. It converts various sensor inputs into optional factors within a single optimization framework, allowing it to share state representation, marginalization, and calibration logic across different sensor combinations. This allows the system to be highly configurable, supporting multiple fusion modes including:

• WIO: Wheel-Inertial Odometry
• VIO: Visual-Inertial Odometry
• LIO: LiDAR-Inertial Odometry
• LVIO: LiDAR-Visual-Inertial Odometry
• Optional integration of wheel encoders and GNSS.

Who it’s for

It is designed for developers and researchers working on autonomous platforms, including wheeled vehicles, legged robots, and aerial UAVs, who need a resilient SLAM system that can adapt to various sensor suites and environmental challenges.

Highlights

• Multi-Platform Support: Validated on wheeled, legged, and aerial platforms.
• High Configurability: A single framework that can be toggled between different sensor fusion modes (WIO, VIO, LIO, LVIO).
• Resilience: Maintains localization availability during sensor degradation and spatiotemporal perturbations.
• Dual Middleware: Provides runtimes for both ROS1 (Noetic) and ROS2 (Humble).
• Omnidirectional Support: Includes Ultra-Fusion-Omni (UFO) for 360° multi-camera LVIO setups.