Semantic Occupancy Mapping

Semantic occupancy mapping estimates which parts of 3D space are occupied, free, or unseen and assigns semantic labels, giving robots and autonomous vehicles a shared scene representation for prediction and action. Methods range from camera- and LiDAR-to-voxel scene completion and temporal sensor fusion to sparse voxel or octree maps, Gaussian primitives, and ray-based models; newer systems add open-vocabulary vision-language features and persistent map memory. Driving benchmarks such as nuScenes/Occ3D and SemanticKITTI test voxel accuracy and forecasting, while ScanNet and embodied indoor benchmarks test online mapping and transfer across rooms. A central challenge is balancing detailed, temporally consistent geometry and semantics against compute, annotation, and memory costs, especially under occlusion and when maps must generalize to unfamiliar environments.