ROAMing: Real-Time Occlusion-Aware 4D Surfel Mapping via Neural Scene Flow

As robots enter everyday environments, modeling dynamic elements in the scene is particularly important. Existing approaches face several limitations: they typically discard dynamic objects as outliers, rely on explicit segmentation masks to reconstruct them, fail to maintain geometry during temporary occlusions, or fail to achieve real-time performance. In this work, we propose a real-time mapping system that naturally handles scene dynamics. Our approach maintains a lightweight surfel map and uses a learned neural flow module to predict the continuous 4D motion of both visible and non-visible surfels. By warping the map forward in time before fusing incoming frames, we track deformations and recover temporarily occluded geometry. Extensive comparisons show state-of-the-art dynamic mapping and tracking accuracy competitive with offline trackers, while running causally in real time. The fast variant of our system updates the map in 56ms.