FAST-LIVO2

Fast, direct LiDAR-inertial-visual odometry from HKU-MARS, published in IEEE T-RO

License

GPL-2.0

Overview

FAST-LIVO2, from the University of Hong Kong's MaRS Lab (the group behind FAST-LIO), is a fast, direct LiDAR-inertial-visual odometry system published in IEEE Transactions on Robotics. It tightly fuses all three sensor streams in an error-state iterated Kalman filter using direct methods — raw lidar points and image photometrics, no feature extraction — achieving robust real-time estimation in environments where lidar-only or visual-only systems degenerate, and it has been demonstrated onboard UAVs for aggressive autonomous flight.

The system's design principle is directness: lidar measurements contribute plane-to-point residuals on raw points, and the camera contributes photometric errors on map points with online-estimated exposure compensation — skipping feature detection entirely keeps latency low and robustness high in texture-poor or structure-poor scenes where one modality fails. A unified voxel map holds geometry and appearance together; the sequential update order (lidar, then vision) inside one error-state iterated Kalman filter lets each sensor correct the state at its own rate. The result runs in real time on embedded computers, handles solid-state lidars with small FoV (Livox), and degrades gracefully: lidar keeps the system alive in visual darkness, vision carries it through geometric degeneracy like long corridors. The repo ships ROS integration, datasets and configurations, with the companion FAST-LIVO2 hardware/software stack used in the lab's UAV work. Alongside FAST-LIO2, it anchors the HKU-MARS family that has become the reference open lineage for tightly-coupled lidar fusion odometry.

Why it matters

FAST-LIVO2 is the flagship of the most influential open lidar-fusion lineage in robotics (HKU-MARS): a T-RO-published system that made three-sensor direct fusion practical on embedded hardware, and the reference point every new LIVO paper measures against.

Problem it solves

Lidar odometry dies in geometric degeneracy; visual odometry dies in darkness and texture-poor scenes. FAST-LIVO2 fuses lidar, camera and IMU tightly enough that either failure mode is covered by the other sensor — in real time on small onboard computers.

Key takeaways

• Tightly-coupled LiDAR-inertial-visual fusion in one error-state iterated Kalman filter
• Direct methods on raw points and photometrics — no feature extraction
• Unified voxel map holding geometry and appearance with online exposure estimation
• Real-time on embedded computers; proven onboard UAVs for autonomous flight
• Published in IEEE T-RO from the FAST-LIO family lab (HKU-MARS)

Who it is for

• SLAM and state-estimation researchers
• UAV and mobile-robot autonomy engineers
• Teams fusing solid-state lidars with cameras
• Graduate students in multi-sensor fusion

Limitations

• GPL-2.0 license restricts closed-source use
• Careful sensor synchronization and calibration are prerequisites
• Tuned primarily around Livox-class solid-state lidars