TRGS-SLAM: IMU-Aided Gaussian Splatting SLAM for Blurry, Rolling Shutter, and Noisy Thermal Images

Researchers Spencer Carmichael and Katherine A. Skinner have introduced TRGS-SLAM, a breakthrough system that allows robots to see and navigate in conditions impossible for standard cameras. The system combines 3D Gaussian Splatting (3DGS)—a cutting-edge neural rendering technique—with data from an inertial measurement unit (IMU) to create accurate 3D maps from the notoriously difficult output of uncooled microbolometer thermal cameras. These affordable thermal sensors are crucial for robotics but produce images plagued by motion blur, rolling shutter distortion, and fixed pattern noise, which cause all existing SLAM (Simultaneous Localization and Mapping) systems to fail.

TRGS-SLAM overcomes these challenges with several key innovations, including a model-aware 3DGS rendering method specifically designed for thermal data and a two-stage IMU loss function for B-spline trajectory optimization. The result is a system that demonstrates robust tracking on real-world, high-speed, noisy thermal data where other methods completely break down. Furthermore, the team showed that by refining the SLAM results offline, they could even perform thermal image restoration at a quality competitive with prior methods that required perfect, ground-truth positioning data.

This advancement unlocks the practical use of thermal vision for autonomous systems. Thermal cameras provide a passive, low-power solution that works in total darkness, is unaffected by blinding lights or shadows, and can see through obscurants like fog, dust, and smoke. TRGS-SLAM effectively makes this sensor modality viable for reliable robot navigation, opening the door to drones that can fly in a burning building or ground vehicles that can operate in a dust-filled mine.