SpotlessGS relights 3D Gaussian Splatting for robots in dark environments

Robots navigating dark or poorly lit spaces often rely on onboard lights, which create uneven illumination that degrades perception. Previous 2D image enhancement methods lacked reliable supervision and broke multi-view geometric consistency. SpotlessGS, developed by Liang Hong, Jiaxin Wei, Simon Schaefer, Stefan Leutenegger, and Jaehyung Jung, builds on Dark Gaussian Splatting to create a more accurate relightable 3D reconstruction framework. The key advance is eliminating explicit light parameter calibration: the system jointly optimizes lighting within the Gaussian Splatting optimization itself, adapting to dynamic illumination without external setup.

SpotlessGS models spatially varying residual and ambient lighting using low-frequency spherical harmonics, while a MLP-based BRDF captures non-Lambertian reflectance for realistic surface rendering. On synthetic and real-world datasets, the method mitigates illumination artifacts and improves rendering quality over prior approaches. The team also validated its benefits through a downstream robotic perception task, showing tangible gains. Accepted to IROS 2026, SpotlessGS offers a practical path to robust 3D scene understanding for robots operating under challenging, dynamic lighting conditions.