In plain words
A robot needs a map that can keep changing as it explores, while still making good use of everything it has seen once exploration ends. VEOcc++ is a mapping framework that keeps a shared map of features as the robot moves, then revisits that map afterward.
On a sequential mapping test, its online map scored higher than the compared methods. A later refinement step improved the map using the same observations, without adding new areas.
The research
Can a robot keep a useful map while exploring and still improve it after the journey, using only observations already collected? Online updates must incorporate views in sequence, but a completed map can use broader scene context. The question is whether one representation can support both stages.
VEOcc++ keeps voxel features in a shared map during exploration, then uses cached evidence and spatial context to refine the completed map.
Can a robot keep a useful map while exploring and improve it after the journey, using observations already collected? VEOcc++ retains voxel features in a shared map, then uses cached local evidence and broader spatial processing to refine the completed map.
On EmbodiedOcc-ScanNet, the online variant reports the strongest occupancy and semantic scores among the compared methods. Offline refinement improves those reported scores without new observations or map expansion, within the evaluated setting.
The benchmark results concern sequential mapping on EmbodiedOcc-ScanNet. The real-world demonstrations show recognizable maps in two indoor environments, but report no numerical accuracy measure there.