VEOcc: Voxel-Centric Online Semantic Occupancy Prediction For Embodied Scene Understanding
Researchers have introduced VEOcc, a novel voxel-centric framework designed for online 3D occupancy prediction and mapping. This system incrementally builds dense spatial representations on the fly, overcoming limitations of previous Gaussian-centric methods that struggled with boundary fidelity and required predefined scene-size priors. VEOcc employs a recursive perception-and-assimilation paradigm, enabling open-ended map expansion without initial scale estimation, and utilizes a Spatio-Temporal-Aware Online Update Strategy for robust temporal observation aggregation. AI
IMPACT Establishes new state-of-the-art in embodied scene understanding, offering a more efficient solution for autonomous exploration.