Online 3D Gaussian Splatting (3DGS) is a promising route for robots building real-time scene representations, but real-world motion causes blurred images that corrupt geometry and drift camera poses. In a new arXiv preprint, Jung et al. present RoMo-GS, a system that treats motion blur as part of the mapping process rather than a nuisance to remove.

Prior motion blur-aware 3DGS methods average N discrete sub-frames, which is expensive. RoMo-GS instead analytically integrates each projected Gaussian over the exposure interval, reducing complexity to O(1) per Gaussian. The system also bounds Gaussian growth with a Hessian-based pruning score, and jointly refines pose, velocity, and Gaussians using Adam with an online Levenberg-Marquardt step over the Gaussians.

Across multiple public datasets and a self-collected dataset, RoMo-GS matches or exceeds offline deblurring 3DGS methods in rendering quality while running 15–150x faster. Compared with an online 3DGS baseline, it operates at competitive quality with about 20% fewer Gaussians. The authors position RoMo-GS as a practical step toward deploying 3DGS on real robots under rapid maneuvers.