Robot mapping software has a ghost problem — and a new filter called MoPe is designed to fix it.
Current monocular Gaussian Splatting SLAM systems build high-fidelity 3D maps from single-camera feeds, but they evaluate each frame in isolation. When a pedestrian pauses or steps behind a pillar, the system may decide that region looks static and bake the person into the map permanently — leaving a ghostly duplicate when they move again. Researchers behind MoPe argue the flaw is architectural: dynamic-ness is a property of motion history, not of any single frame. Their fix, called Motion Permanence, carries an object's "dynamic identity" forward through time using geometry-consistent SE(3) warping and Bayesian log-odds updates, so a momentarily-still person doesn't get mistaken for furniture.
The practical stakes are higher than clean visuals. Localization and navigation decisions downstream depend on the map being accurate, and a map that absorbs pedestrians as static geometry is a liability in any real-world deployment. On standard benchmarks — Wild-SLAM, Bonn, and TUM sequences — MoPe improved tracking robustness and cut residual ghosting, with the biggest gains on scenes involving moving humans.
Gaussian Splatting SLAM is still young enough that most competitive comparisons are between academic preprints rather than shipping products, but the memoryless-frame problem MoPe targets is a known weak spot the field hasn't cleanly solved. Whether the SE(3) warping overhead is acceptable on edge hardware is the question nobody's answered yet.