Real-time Gaussian Splatting SLAM Family

Lab. of Artificial Intelligence and Robotics (LAIR)

Page will be updated soon.

We present a family of real-time Gaussian SLAM systems across RGB-DRGB-D, RGBmonocular RGB, and LiDARLiDAR sensing. Our goal is to build efficient 3D Gaussian spatial representations for robot localization, mapping, and future sensor-agnostic spatial intelligence RGB-D. These are not isolated systems, they are three modality-specific instances of a broader research direction: real-time Gaussian spatial representations for robotics.

Common Design Principles

Real-time Gaussian scene representation for robotics

  • Tight coupling between tracking and mapping Conversion path Sync Map
  • Sensor-specific observation models over a shared spatial representation Diversity
  • Online map optimization and scalable map management Map Rebase edit
  • Toward sensor-agnostic Gaussian SLAM and map handover Multi-robot Sync Robot intelligence

Papers

Real-time Gaussian SLAM papers

RGB

GSO-SLAM: Bidirectionally Coupled Gaussian Splatting and Direct Visual Odometry

RAL 2026

Jiung Yeon*, Seongbo Ha*, Hyeonwoo Yu

Monocular Gaussian SLAM through bidirectional coupling between direct visual odometry and Gaussian mapping.

* Equal Contribution

GSO-SLAM

System comparison

System Input Main Challenge Tracking Cue Core Contribution
RGBD GS-ICP SLAM RGB-D Dense real-time RGB-D mapping Depth / G-ICP High-speed coupled Gaussian SLAM
GSO-SLAM RGB Monocular scale and geometry Direct visual odometry Bidirectional pose-map coupling
LiDARGS-SLAM LiDAR Sparse range-only large-scale mapping G-ICP covariance Covariance-driven map control

Future goal

Toward Sensor-Agnostic Gaussian SLAM
These systems cover complementary sensing regimes: RGB-DRGB-D, RGBmonocular RGB, and LiDARLiDAR sensing.
Our next goal is to study how Gaussian maps can be shared, handed over, and continuously optimized across heterogeneous sensors.